Methods and processing systems for use with assets of an electric power system and methods of providing physical information machine learning models

Abstract

To process measurement results acquired using measurement instruments (41, 42), a processing system (50) is operable to perform a physics information data-driven processing technique to determine at least one parameter related to an asset (20) of an electric power system (10). The processing system (50) is operable to perform a pre-processing (70) of the measurement results to generate input data, wherein the pre-processing (70) comprises identifying a subset of measurement locations for which the measurement results are reliable for input to the data-driven processing technique, and using measurement results obtained at the identified subset to generate the input data. The processing system (50) is operable to estimate local heat losses using the data-driven processing technique.

Description

Technical Field

[0001] Embodiments of the present invention relate to processing systems and methods for controlling and / or monitoring power systems. Specifically, embodiments of the present invention relate to processing systems and methods capable of processing measurement results obtained at or near power system assets and triggering actions in response to such processing. Background Technology

[0002] Modern power systems include several complex assets, such as transformers or shunt reactors. Failure or malfunction of such assets can have potentially catastrophic consequences. Monitoring such assets can be a considerable challenge. For illustration, some assets may have enclosures, including walls. Localized losses in such assets can be difficult to monitor, partly due to the structural components of the enclosure walls. Therefore, there is a risk that potential problems associated with localized losses in, for example, such complex assets, may go unnoticed for unintended extended periods.

[0003] US 2020 / 0293594 A1 discloses a technique for modeling objects using differential equations. F. Bragone et al., “Physics-informed neural networks for modelling power transformer's dynamic thermal behavior,” *Electric Power Systems Research*, Vol. 211, October 2022, 108447, Elsevier (2022) discloses the use of Physical Information Neural Networks (PINN) for thermal modeling of power transformers. C. Moya et al., “DeepONet-Grid-UQ: A Trustworthy Deep Operator Framework for Predicting the Power Grid's Post-Fault Trajectories,” arXiv:2202.07176 discloses a data-driven method for predicting post-fault trajectories in power systems.

[0004] The processing techniques discussed in the preceding literature provide data-driven technologies suitable for assets. These techniques are beneficial because they reduce the risk of human error. However, there is still a need for enhanced techniques that allow the detection of localized losses, malfunctions, or other potentially related operational conditions. For illustration, it is desirable to provide techniques that can infer conditions at locations where measurements obtained from instrument locations, such as outside the asset enclosure walls, cannot reliably reflect thermal diffusion processes within the asset due to structural components of the enclosure walls. Typically, there is a need for technologies suitable for use in conjunction with assets in power systems that provide enhanced robustness. Summary of the Invention

[0005] The object of this invention is to provide a method and / or processing system for implementing enhanced techniques for processing measurement results obtained from assets of a power system. In particular, there is a need for such a method and / or processing system that can determine one or more parameters associated with the asset and provide enhanced robustness.

[0006] According to exemplary embodiments, the method and processing system described in the independent claims are provided. Dependent claims define preferred or advantageous embodiments.

[0007] According to an exemplary embodiment, a method and processing system are provided capable of operating a data-driven processing technique to process measurement results obtained on or near an asset in a power system. The data-driven processing technique may be a physical information data-driven processing technique capable of determining one or more parameters of the asset, wherein knowledge of differential equations (such as partial differential equations (PDEs), e.g., linear PDEs) relating various parameters to each other is embedded in the physical information data-driven processing technique. The input data for the data-driven processing technique is generated based on the measurement results. Generating the input data may include identifying measurement locations where measurement results were obtained but are unreliable (in this sense, the measurement results do not conform to the physical knowledge embedded in the data-driven processing technique).

[0008] Enhanced robustness is achieved by generating input data for data-driven processing techniques in a manner that takes into account the physical processes (such as thermal diffusion) embedded in the data-driven processing techniques at which measurement locations the measurement results are determined.

[0009] Output is generated based on one or more defined parameters. The output may include warnings, alerts, alarms, or control commands applied to power system equipment.

[0010] Data-driven processing techniques can include physical information data-driven processing techniques capable of determining one or more parameters of an asset. These techniques embed knowledge of physical processes (e.g., thermal diffusion) described by partial differential equations (PDEs) (e.g., linear PDEs). The physical information data-driven processing techniques can be used to determine a first parameter and a second parameter based on both spatial location and time on the asset. The first and second parameters can be correlated with each other via a PDE (particularly a linear PDE). The PDE can depend on both the time derivative of the second parameter and the first parameter's first-order partial time derivative and second-order partial spatial derivative (e.g., a Laplace operator applied to the first parameter). The PDE can be a diffusion equation, such as a thermal diffusion equation.

[0011] Thus, data-driven processing technology embeds physical relationships between several parameters, and these parameters are interconnected through these physical relationships.

[0012] The example is heat and temperature correlated with each other through the thermal diffusion equation. A more robust determination of the parameters can be obtained by determining these first and second parameters based on time and location, and by using a PDE that correlates the first and second parameters. Robust parameter determination is achieved by combining the use of physical information data-driven processing techniques with the preprocessing of measurement locations where the measurement results do not reflect the physical processes embedded in the physical information data-driven processing techniques.

[0013] Data-driven processing technology can be used to determine heat in a spatially resolved, optionally spatiotemporally resolved manner, manner using temperature measurements obtained from multiple measurement locations on the walls of an asset (e.g., transformer wall or shunt reactor wall).

[0014] Therefore, the processing system and processing method can be used to identify local losses (such as heat losses identified in a spatiotemporal resolution manner).

[0015] Data-driven processing technology can be used to determine the temperature of areas where reliable temperature measurements cannot be obtained, for example, due to structural components of the enclosure (such as reinforcements).

[0016] This allows for the acquisition of additional information that is not available using traditional technologies, which can be used to generate outputs such as alerts, warnings, alarms, or other data that triggers actions in the power system (such as control commands).

[0017] The processing system and methods can be operated to create a thermodynamic baseline during normal operation of the asset. The processing system and methods can also be operated to identify deviations from the baseline and generate outputs based on these deviations.

[0018] Thus, the processing system and processing method can be operated to generate outputs (such as warnings, alerts, alarms, or control commands applied to power system equipment) based on automatically detected anomalies (in the sense that they deviate from baseline thermodynamic behavior).

[0019] A method for use in association with an asset of a power system, according to an embodiment, includes receiving data at at least one interface of a processing system, the data including measurement results at or near the asset at several different measurement locations. The method includes determining at least one parameter of the asset by at least one processing circuit of the processing system, including generating input data for physical information data-driven processing techniques based at least on the measurement results, including identifying a subset of measurement locations where the measurement results are reliable and using the measurement results obtained at the identified subset to generate the input data, and applying physical information data-driven processing techniques to the input data to determine the at least one parameter. The method includes generating an output by the processing system based on the at least one parameter.

[0020] Therefore, the input data for generating physical information data-driven processing techniques includes determining measurement results that are reliable for processing by these techniques. Measurement results that do not reflect the physical processes embedded in the physical information data-driven processing techniques can be removed; that is, they are not used to generate the input data for these techniques. This enhances robustness and accuracy.

[0021] Measurement results may include spatially resolved temperature measurements. Identifying a subset of measurement locations may include identifying spatial regions where the spatially resolved measurements do not indicate physical processes (such as thermal diffusion) embedded in physical information data-driven processing techniques.

[0022] Therefore, input data for physical information data-driven processing can be generated based on measurements obtained at a measurement location (e.g., at a measurement location on the wall of an asset) that reflect a physical process (e.g., thermal diffusion) embedded in the physical information data-driven processing technique, while being generated independently of other measurements obtained at other measurement locations (e.g., at other measurement locations on the wall of an asset) that do not reflect the physical process (e.g., thermal diffusion). This enhances robustness.

[0023] Measurement results that do not indicate physical processes embedded in physical information data-driven processing techniques may include measurements obtained at locations where structural components of an asset attenuate thermal radiation and / or act as radiators.

[0024] This mitigates the risk of incorrect asset parameter determination. The technique also mitigates the risk that thermal measurements that do not indicate thermal diffusion (for which physical information data-driven processing techniques have been trained) may reduce the reliability of the output.

[0025] Identifying a subset of measurement locations may include at least one of the following operations: performing image segmentation based on a temperature gradient determined from temperature measurements and / or based on asset construction data.

[0026] Therefore, it is possible to identify areas (e.g., areas on the walls of an asset) where the measurement results do not reflect physical processes (e.g., thermal diffusion) embedded in physical information data-driven processing techniques.

[0027] Segmentation can be based on the spatial derivative of the measurement results.

[0028] Therefore, the boundaries of regions (e.g., regions on the walls of an asset's enclosure) can be identified based on measurement results, where the first spatial derivative of the measurement results exhibits discontinuity, indicating that the thermal measurements at that region are not controlled by thermal diffusion processes. The input data for physical information data-driven processing techniques can be generated independently of regions with boundaries exhibiting discontinuities in the first spatial derivative of the measurement results.

[0029] Spatial-resolved temperature measurement results may include pixel or voxel values ​​of at least one thermal image of at least one wall of the asset's enclosure.

[0030] Therefore, asset parameters can be determined based on thermal images of one or more of the side walls and / or lid of the enclosure. These measurements can be obtained without the need for instruments inside the enclosure.

[0031] At least one parameter can indicate electrical power loss.

[0032] Therefore, this method can be used to identify localized damage in assets.

[0033] Physical information data-driven processing techniques may include at least one physical information machine learning (PIML) model.

[0034] Therefore, data-driven processing techniques can be used to determine at least one parameter by embedding at least one physical relation (such as a partial differential equation (PDE), particularly a linear PDE or other differential equation) that conforms to at least one parameter.

[0035] The at least one PIML model can be trained using a loss function that includes terms that depend on the residuals of the PDE.

[0036] Therefore, the PIML model (using a loss function term that depends on the PDE residuals) provides an output that conforms to the physical relationships embedded in the PIML model. This necessitates the identification of physically plausible solutions through data-driven processing techniques.

[0037] The at least one PIML model may have an output terminal capable of operating to output a first parameter (e.g., temperature) and a second parameter (e.g., heat) for each of a plurality of spatial locations. The at least one PIML model may also be operable to provide these values ​​at the output terminal for each of a plurality of time periods.

[0038] Thus, the at least one PIML model is operable to determine at least one parameter in a spatially resolved manner, and optionally in a spatiotemporally resolved manner. This facilitates the spatially resolved determination of one or more parameters such that they conform to physical relationships (e.g., heat diffusion equations) embedded in the at least one PIML model.

[0039] The at least one PIML model may have inputs that are operable to receive the spatial coordinates of the asset, time, measurement results (such as thermal measurement results), and at least one operating condition (such as load current).

[0040] Therefore, for cases where at least one parameter depends on operating conditions, it facilitates the determination of at least one parameter. For illustration, it facilitates the determination of heat and / or temperature in a spatially or spatiotemporally resolved manner, wherein the PIML model receives asset load current or load power at the input, in particular.

[0041] The at least one PIML model may include: a first plurality of neural network (NN) layers having a first input layer; a second plurality of NN layers operable to process a first output of the first plurality of NN layers and having a second output for providing a first parameter (e.g., temperature); and a third plurality of NN layers operable to process the first output of the first plurality of NN layers and having a third output for providing a second parameter (e.g., heat) related to the first parameter through a physical relationship (e.g., thermal diffusion) embedded in the at least one PIML model (e.g., by using a PIML model trained with a loss function that depends on the residuals of a partial differential equation for thermal diffusion).

[0042] This facilitates the training of at least one PIML model. The PIML model is determined by several hidden layers (a first plurality of NN layers) shared by several parameters that are physically related to each other.

[0043] The at least one PIML model can be operated to provide a first parameter and a second parameter, wherein the at least one PIML model embeds a PDE that depends on the first parameter and the second parameter.

[0044] Thus, the at least one PIML model is capable of determining several parameters that are related to each other through physical relationships (such as the thermal diffusion equation). This enhances robustness and reliability. In particular, the at least one PIML model mandates finding physically plausible solutions for the first and second parameters using a PDE embedded in the PIML model. As those skilled in the art will understand, embedding the PDE into the PIML model can be achieved by training the PIML model with a loss function that includes a loss function term dependent on the PDE residuals.

[0045] A PDE can be a linear PDE. A PDE can be a PDE that includes a first-order partial time derivative of a first parameter and at least two second-order partial spatial derivatives (such as 2D or 3D Laplace operators) and / or a second parameter.

[0046] Therefore, the PIML model can be used to determine several parameters that are related to each other through diffusion-type PDEs.

[0047] PDEs can include thermal diffusion relationships.

[0048] Therefore, this method is capable of determining temperature and / or heat in a spatially resolved manner, and optionally in a spatiotemporally resolved manner. This is useful for identifying asset conditions associated with localized heat and temperature increases, such as situations where localized losses may occur.

[0049] At least one PIML model can be embedded in the thermal diffusion partial differential equation (PDE).

[0050] Therefore, this method can be operated to determine temperature and / or heat in a manner consistent with the physical process of thermal diffusion.

[0051] The at least one PIML model can be operated to receive spatially resolved measurements of a first parameter (e.g., thermal measurements captured using a thermal camera) at its input and to provide spatially resolved values ​​of both the first parameter (e.g., temperature) and the second parameter (e.g., heat) at its output. The at least one PIML model can be operated such that its input does not receive measurements of the second parameter. The at least one PIML model can be operated such that its input receives measurements of at least one operating parameter (e.g., load current and / or load power) that affects the second parameter.

[0052] Thus, the method is capable of determining the first parameter (e.g., temperature) and the second parameter (e.g., heat) in a manner that takes into account the physical process that associates the first and second parameters, without requiring measurements of the second parameter to be available as input data for generating the at least one PIML model.

[0053] The at least one PIML model includes at least one output layer that is operable to enforce the positivity of temperature and heat determined by the at least one PIML model.

[0054] This enhances robustness.

[0055] The at least one output layer may include an output layer that determines the modulus (i.e., the absolute value).

[0056] This enhances robustness.

[0057] The at least one PIML model may be a PIML model trained using training data captured on the asset or on another asset having the same or similar structural characteristics as the asset. The training data may include spatiotemporally resolved temperature measurements.

[0058] Therefore, thermal diffusion processes and / or localized losses in assets can be determined in a data-driven manner, while the determined parameters conform to the underlying physical processes (e.g., thermal diffusion).

[0059] The method further includes training the at least one physical information data-driven processing technique by a processing system or by a computing system separate from the processing system.

[0060] Therefore, this at least one physical information data-driven processing technique is configured to be used during inference.

[0061] Training can include historical awareness training.

[0062] Therefore, techniques that allow for efficient training can be used to obtain enhanced robustness of the acquired parameters determined during inference.

[0063] History-aware training can be performed using a loss function (e.g., a loss function to be maximized or minimized during training), where the value of the loss function at any point during training depends on the value of the loss function at at least one previous point during training.

[0064] Therefore, techniques that allow for efficient training can be used to obtain enhanced robustness of the acquired parameters determined during inference.

[0065] History-aware training can be performed using a loss function that is to be maximized or minimized. The loss function may include a weighting factor that depends on the training loss at past training moments and / or introduces temporal gradient regularization and / or penalizes the training loss based on the derivative of the loss function.

[0066] Therefore, techniques that allow for efficient training can be used to obtain enhanced robustness of the acquired parameters determined during inference. It has been unexpectedly found that, in several previous instances, weight functions that depend on the time derivative of the loss function offer the advantages of both robustness and training efficiency.

[0067] The method may further include performing validation on trained physical information data-driven processing techniques, wherein the validation is performed at least based on spatially resolved calorific values ​​obtained by performing simulations of the asset's thermodynamics.

[0068] Thus, physical information data-driven processing techniques can be configured to determine temperature and heat in a spatially resolved manner (optionally in a spatiotemporally resolved manner) during inference, without requiring spatially resolved heat measurements for verification.

[0069] The method may further include interface connecting the processing system or the computing system to an asset simulator that simulates the thermal behavior of the asset, in order to perform model verification on the at least one PIML model.

[0070] Thus, physical information data-driven processing techniques can be configured to determine temperature and heat in a spatially resolved manner (optionally in a spatiotemporally resolved manner) during inference, without requiring spatially resolved heat measurements for verification.

[0071] Assets may include containers with walls. Data may include thermal measurements obtained at and / or within the container walls.

[0072] Therefore, this method is capable of determining temperature and / or heat in a spatially resolved manner, and optionally in a spatiotemporally resolved manner. This is useful for identifying asset conditions associated with localized heat and temperature increases, such as situations where localized losses may occur.

[0073] At least one parameter may include temperature and / or heat.

[0074] Therefore, this method is capable of determining temperature and / or heat in a spatially resolved manner, and optionally in a spatiotemporally resolved manner. This is useful for identifying asset conditions associated with localized heat and temperature increases, such as situations where localized losses may occur.

[0075] The enclosure wall may have a bottom surface, several side surfaces, and a lid. The data may include thermal measurements obtained for at least two or more (optionally all) of the side surfaces. The data may also include thermal measurements obtained for the lid.

[0076] Therefore, a more comprehensive determination of possible local losses can be obtained using data that can be easily captured at various parts of the tank wall (e.g., transformers or shunt reactors) than that based on measurements performed, for example, at only a single tank wall. This is beneficial for determining the thermal behavior of several tank sidewalls along the tank sidewalls and optionally the ground cover.

[0077] The data may include measurements obtained on a portion of the tank wall near the asset's high-pressure side and another portion of the tank wall near the asset's low-pressure side.

[0078] This allows for more comprehensive thermodynamic modeling of assets, such as identifying localized losses.

[0079] Preprocessing the measurement results may include applying coordinate transformations to transform the coordinates of the pixel or voxel values ​​included in the measurement results to the world coordinate system of the asset.

[0080] This allows for compensation for changes in the attitude and / or position of measuring instruments (such as thermal cameras or other thermal sensors) relative to the chamber wall, resulting in increased robustness and reliability.

[0081] Alternatively or additionally, preprocessing may include segmenting at least one array of pixel or voxel values ​​included in the measurement results to identify unreliable partitions of the measurement results.

[0082] Therefore, through preprocessing, measurements in areas where the values ​​of data-driven processing techniques are reduced (e.g., due to the presence of structural box components, such as box reinforcements, or box bars) can be disregarded. This can include cropping the pixel or voxel array. This has the additional technical advantage of identifying areas from the thermal measurement data itself that could render the results unreliable, for example, due to the presence of box reinforcements. Cropping reduces the risk of artifacts in areas where box reinforcements would otherwise cause discontinuities in temperature measurements (due to the added thermal shielding of the box reinforcements).

[0083] Alternatively or additionally, preprocessing may include identifying pixel or voxel values ​​of partitions of the asset’s structural components that reduce the reliability of measurement results, including using interpolation, extrapolation, or regression techniques.

[0084] Therefore, for areas where the measurement results of data-driven processing techniques decrease (e.g., due to the presence of structural box components, such as box reinforcements, such as box bars), the measurement results can be replaced by more reliable alternative data obtained using interpolation, extrapolation, or regression techniques.

[0085] Measurements may include spatially resolved thermal measurements.

[0086] Therefore, potential problems in asset operation can be identified by using, for example, the spatiotemporal temperature and / or changes on the walls of the asset, which can be used to trigger warning, alarm, or control actions on power system equipment based on at least one parameter.

[0087] Spatially resolvable thermal measurement results can be received from a measuring instrument, which may include one, several, or all of the following: at least one thermal camera, at least one thermocouple, and at least one fiber optic sensor capable of operating for sensing temperature.

[0088] Therefore, the input for physical information data-driven processing technology can be generated based on these thermal measurement results.

[0089] Data-driven processing techniques can be used to model the thermodynamics of an asset in a spatially resolved manner along at least two spatial directions, taking into account thermal diffusion.

[0090] Therefore, thermal effects can be embedded in data-driven processing techniques (e.g., via a PIML model trained with a loss function that depends on the residuals of the heat diffusion equation). The identification of localized losses can be facilitated by determining the location of temperature and / or heat based on the planes along at least one or more sidewalls of the asset box. Additional information useful for determining which output to generate is obtained. For example, based on the location of temperature accumulation identified using a 2D heat diffusion model embedded in the data-driven processing technique, it can be determined which of the following actions should be performed: generating a warning; generating an alarm; triggering control commands on key system equipment affecting the power system.

[0091] The output may include the output of the operator interface of the power system.

[0092] Therefore, the determined at least one parameter can be used for warning, alert, and / or alarm generation. Thus, this method can be an operational method for a warning, alert, and / or alarm generation system, wherein warnings, alerts, or warnings are selectively issued based on the determined at least one parameter.

[0093] Alternatively or additionally, the output may include an output that triggers at least one control action within the power system. Control actions may affect major system equipment, for example, by changing asset load current and / or asset load power, by operating protective devices (such as circuit breakers or switches), by changing tap changer positions, or by other control operations.

[0094] Assets can have box walls, where input data for data-driven processing techniques can be generated based on measurements obtained from the box walls of the asset.

[0095] Therefore, measurement results that can be obtained from the outside of the chamber wall can be used in the processing method. This reduces the risk of measuring instruments inside the chamber introducing potential problems.

[0096] The walls of the enclosure may contain insulating fluids, such as insulating oil.

[0097] Therefore, the technology disclosed herein can be used in connection with oil-insulated assets.

[0098] A method for controlling a power system according to an embodiment includes: a processing system performing a method of any aspect or embodiment disclosed herein to (e.g., based on the location of an asset in an asset box) identify local losses in assets of the power system; and the processing system or a control system operatively coupled to the processing system performing at least one control operation based on the output.

[0099] Therefore, the technology disclosed in this paper is used to automatically perform control operations.

[0100] One method for providing a PIML model includes: obtaining training data by a computing system, the training data being based on measurements captured on the box walls of an asset. The method includes training the PIML model by the computing system using the training data, wherein the training includes historically-aware training where the loss function at training time depends on a loss function at at least one past training time.

[0101] This enhances the convergence of training and improves the accuracy of the output (i.e., the output of at least one parameter). History-aware training considers the dynamics of physical processes (such as diffusion processes), making the accuracy of dynamics susceptible to past estimation errors. Therefore, the PIML model is trained in a way that takes into account the fact that physical relationships (such as diffusion processes) are embedded in the PIML model (e.g., penalizing the residuals of the PDEs representing physical relationships during training). This is particularly important for setting up a PIML model that is used to predict two parameters that are correlated with each other through diffusion processes (such as heat and temperature correlated through thermal diffusion processes) in a spatiotemporally resolved manner.

[0102] History-aware training can rely on the time derivative of the loss function.

[0103] This improves training convergence and enhances the accuracy of the output (i.e., the output of at least one parameter). History-aware training takes into account past variations in the loss function.

[0104] History-aware training can be performed using a loss function that is to be maximized or minimized. The loss function may include a weighting factor that depends on the training loss at past training moments and / or introduces temporal gradient regularization and / or penalizes the training loss based on the derivative of the loss function.

[0105] Therefore, techniques that allow for efficient training can be used to obtain enhanced robustness of the acquired parameters determined during inference. It has been unexpectedly found that, in several previous instances, weight functions that depend on the time derivative of the loss function offer the advantages of both robustness and training efficiency.

[0106] This method may further include PIML model validation. PIML model validation may include using spatially resolved thermal predictions provided by the PIML model to determine the spatially resolved simulated temperature distribution by a computing system or a simulator connected to the computing system via a communication interface, and comparing the spatially resolved temperature predictions with spatially resolved temperature measurements.

[0107] Therefore, the performance of the PIML model can be evaluated without the need for actual heat measurements, which may be more difficult to obtain than temperature measurements.

[0108] The method may include providing PIML model data obtained during training for processing measurement results as described in any of the preceding claims.

[0109] Thus, at least one PIML model is set up and provided for inferring at least one parameter.

[0110] This method can be executed automatically by a computing system that can be integrated with or communicatively coupled to a processing system that uses the PIML model during inference.

[0111] A PIML model can have a first part that can be operated to provide a first parameter (such as temperature) as a first output, and a second part that can be operated to provide a second parameter (such as heat) as a second output, wherein the first parameter and the second parameter are related to each other by physical relationships (such as thermal diffusion) embedded in the PIML model during training.

[0112] This can enhance the robustness and accuracy gained during reasoning.

[0113] Training a PIML model may include a training phase in which the parameters of the first PIML model in the first part are adjusted while the parameters of the second PIML model in the second part are kept constant, and another training phase in which the parameters of the second PIML model in the second part are adjusted while the parameters of the first PIML model in the first part are kept constant.

[0114] This enhances the convergence and accuracy of the training gained during inference.

[0115] The loss function may include terms that depend on the residuals of the PDEs embedded in the PIML model, such as the residuals of the heat diffusion equation.

[0116] Thus, the PIML model is trained to provide spatiotemporal values ​​of parameters (such as temperature and / or heat) that conform to the PDEs embedded in the PIML model.

[0117] The method of providing a PIML model can be performed using measurement results, which are captured on assets that have the same or similar construction as the assets for which the PIML model is to be used during inference.

[0118] Therefore, the measurements captured during the factory testing phase can be used to train the PIML model.

[0119] The method of providing a PIML model can be performed using measurement results, which are captured on the assets for which the PIML model is used during inference.

[0120] Therefore, measurements captured, for example, during the factory testing phase, can be used to train the PIML model.

[0121] Measurements used in PIML model training can include thermal measurements captured in a spatiotemporally resolved manner.

[0122] Therefore, the thermodynamics observed, for example, during factory testing, can be used to train the PIML model.

[0123] Measurements used for training may include thermal camera images. Thermal camera images may include a series of thermal camera images, such as at least 50 or at least 100 thermal camera images.

[0124] Therefore, a series of thermal camera images captured in, for example, a factory test setting can be used to train the PIML model.

[0125] Methods for providing PIML models may include performing thermal modeling of assets by a computing system or by a simulator connected to the computing system via a communication interface, in order to establish realistic data of thermal data during PIML model training.

[0126] Therefore, thermal modeling can be used during the training phase to generate heat data that may be difficult to measure. Thermal modeling is not required during inference because the knowledge of the heat diffusion equation is embedded in the PIML model during training.

[0127] Additional features of the PIML model that can be used in the method of providing the PIML model correspond to the features explained by the method for use in relation to assets, as disclosed above and elsewhere in this document.

[0128] Methods for providing a PIML model may include generating training data based on measurement results, wherein generating training data includes identifying a subset of measurement locations where the measurement results are reliable and using the measurement results obtained at the identified subset to generate training data.

[0129] Therefore, generating training data involves determining measurements that are reliable for training the PIML model. Measurements that do not reflect the physical processes embedded in the PIML model can be removed, i.e., not used for training. This enhances robustness.

[0130] Measurement results may include spatially resolved temperature measurements. Identifying a subset of measurement locations may include identifying spatial regions where the spatially resolved measurements do not indicate physical processes (such as thermal diffusion) embedded in physical information data-driven processing techniques.

[0131] Therefore, training can be performed based on measurements obtained at measurement locations (e.g., on the walls of the asset's enclosure, where the temperature indicates thermal diffusion within the asset) that reflect the physical processes (e.g., thermal diffusion) embedded in the PIML model. Measurements obtained at other measurement locations (e.g., on the walls of the asset's enclosure, where the presence of radiators or thermal attenuators means the thermal measurements do not need to indicate thermal diffusion within the asset) that do not reflect the physical processes (e.g., thermal diffusion) are not used for training. This enhances robustness.

[0132] Measurements that do not indicate physical processes embedded in the PIML model may include measurements obtained at locations where structural components of the asset attenuate thermal radiation and / or act as radiators.

[0133] Therefore, the PIML model is trained in a way that enhances accuracy and robustness during inference.

[0134] In the method of providing a PIML model, identifying a subset of measurement locations may include at least one of the following operations: performing image segmentation based on a temperature gradient determined from temperature measurements and / or based on the asset's construction data.

[0135] Therefore, it is possible to identify areas (e.g., areas on the walls of an asset) where the measurement results do not reflect physical processes (e.g., thermal diffusion) embedded in physical information data-driven processing techniques.

[0136] In methods that provide PIML models, segmentation can be based on the spatial derivative of the measurement results.

[0137] Therefore, the boundaries of regions (e.g., regions on the walls of an asset) can be identified based on measurements where the first spatial derivative of the measurement results exhibits discontinuities (indicating that the thermal measurements at that region are not controlled by the thermal diffusion process). Training data can be generated independently of regions with boundaries exhibiting discontinuities in the first spatial derivative of the measurement results.

[0138] In the method of providing a PIML model, the spatially resolved temperature measurement results may include pixel values ​​or voxel values ​​of at least one thermal image of at least one wall of the asset's box.

[0139] Therefore, PIML models can be trained to determine thermodynamic behavior and / or identify localized losses in assets.

[0140] Additional preprocessing steps that can be applied to measurement results to generate training data for use in PIML model training correspond to the features explained by the methods used in relation to assets, as referenced above and elsewhere in this document. These preprocessing steps may specifically include one, several, or all of coordinate transformation, clipping, interpolation, and / or extrapolation.

[0141] The processing system according to an embodiment includes at least one interface operable to receive data, including measurement results from several different measurement locations at or near the asset. The processing system includes at least one processing circuit operable to determine at least one parameter of the asset. The at least one processing circuit operable to generate input data for a physical information data-driven processing technique based at least on the measurement results, including identifying a subset of measurement locations where the measurement results are reliable and using the measurement results obtained at the identified subset to generate the input data. The at least one processing circuit operable to apply a physical information data-driven processing technique to the input data to determine at least one parameter. The processing system is operable to generate an output based on the at least one parameter.

[0142] Therefore, this processing system can operate on the input data used to generate physical information data-driven processing techniques, including determining measurement results that are reliable for processing by these techniques. Measurement results that do not reflect the physical processes embedded in the physical information data-driven processing techniques can be removed; that is, they are not used as input data for generating these techniques. This enhances robustness and accuracy.

[0143] The processing system may include, or may be, an alert and / or alarm generation system capable of generating alerts and / or alarms based on the determination. The processing system may include, or may be a control system capable of performing control actions (e.g., control actions affecting major power system equipment) based on the determination.

[0144] The processing system can operate to perform methods for any aspect or embodiment disclosed herein. The processing system may include, or may be communicatively coupled to, a storage system in which parameters of data-driven processing techniques are stored. The processing system can operate to provide output via at least one interface.

[0145] The power system includes assets, measuring instruments operable to obtain measurement results at or near the assets from several different measurement locations, and, according to one aspect or embodiment, a processing system operable to process data obtained using these measurement results.

[0146] Therefore, the technical effects of the processing systems and methods discussed in this paper can be used to achieve enhanced asset monitoring and / or control of power systems.

[0147] The power system may include a power generation system, a power transmission system, and / or a power distribution system.

[0148] Therefore, the technical effects of the processing systems and methods discussed in this paper can be used to achieve enhanced asset monitoring and / or control in systems where conventional processing techniques have shortcomings.

[0149] The power system may include a transmission network and / or a distribution network.

[0150] Therefore, the technical effects of the processing systems and methods discussed in this paper can be used to achieve enhanced asset monitoring and / or control in systems where conventional processing techniques have shortcomings.

[0151] The power system may include a high-voltage direct current (HVDC) system. The power system may include inverters / converters that incorporate valves based on insulated-gate bipolar transistors (IGBTs) or thyristors.

[0152] Therefore, the technical effects of the processing systems and methods discussed in this paper can be used to achieve enhanced asset monitoring and / or control in systems where conventional processing techniques have shortcomings.

[0153] According to a further embodiment, machine-readable instruction code is provided that, when executed by at least one programmable circuit, causes the at least one programmable circuit to perform a processing method, control method, or operation method according to any aspect or embodiment disclosed herein.

[0154] According to a further embodiment, a non-transitory storage medium is provided that stores machine-readable instruction code, which, when executed by at least one programmable circuit, causes the at least one programmable circuit to perform a processing method, control method, or operation method according to any aspect or embodiment disclosed herein.

[0155] Embodiments of the present invention achieve various effects and advantages. For illustration, the processing systems and methods according to the embodiments provide enhancement techniques that can be used in association with asset monitoring and / or power system control. These processing systems and methods apply preprocessing to measurements obtained on assets and generate input data for physical information data-driven processing techniques based on the identification of measurements indicating physical processes (such as diffusion processes) embedded in physical information data-driven processing techniques.

[0156] These processing systems and methods are operable to determine heat and / or temperature at least along several sidewalls of the asset enclosure in a spatiotemporally resolved manner, taking into account thermal diffusion. This facilitates the identification of localized losses within the asset.

[0157] These processing systems and methods can be used in conjunction with, but are not limited to, the power grid or its subsystems. Attached Figure Description

[0158] Embodiments of the present invention will be described with reference to the accompanying drawings, in which similar or identical reference numerals denote elements having similar or identical configurations and / or functions.

[0159] Figure 1 It is a schematic diagram of a system that includes a processing system.

[0160] Figure 2 It is a schematic representation of the processing performed by the processing system.

[0161] Figure 3 It is a schematic representation of the processing performed by the processing system.

[0162] Figure 4 This is a flowchart of the method.

[0163] Figure 5 The measurement results processed by the processing system are shown.

[0164] Figure 6 The results of the preprocessing operations performed by the processing system are shown.

[0165] Figure 7 The result of another preprocessing operation performed by the processing system is shown.

[0166] Figure 8 It shows Figure 7The amplified portion is used to explain another preprocessing operation performed by the processing system.

[0167] Figure 9 Shown at magnified scale Figure 8 Details.

[0168] Figure 10 Shown at magnified scale Figure 8 Another detail.

[0169] Figure 11 The result of another preprocessing operation performed by the processing system is shown.

[0170] Figure 12 This is a flowchart of the method.

[0171] Figure 13 This is a flowchart of the method.

[0172] Figure 14 It is a schematic representation of a PIML model that can be implemented by a processing system.

[0173] Figure 15 It is a schematic representation of a PIML model that can be implemented by a processing system.

[0174] Figure 16 The spatial variation of the difference between the PIML model and the real data is shown for different implementations of the PIML model.

[0175] Figure 17 This is a flowchart of the method.

[0176] Figure 18 This is a flowchart of the method.

[0177] Figure 19 It is a schematic representation of a system that includes a computing system capable of operating to provide a PIML model for use during inference.

[0178] Figure 20 This is a flowchart of the method.

[0179] Figure 21 This is a flowchart of the method.

[0180] Figure 22 The deviations between the PIML model outputs and real data are shown in several implementations of training a history-aware PIML model and training a non-history-aware PIML model.

[0181] Figure 23 The deviations between the PIML model output and real data from several implementations of training a history-aware PIML model are shown, where the number of past time instances is considered different for various curves.

[0182] Figure 24 This is a flowchart of the method.

[0183] Figure 25 This is a flowchart of the method.

[0184] Figure 26 The results show a comparison between the heat output by the PIML model and the heat determined using simulation tools for various locations on the tank wall.

[0185] Figure 27 It shows in Figure 26 The data is provided by the arrangement of various locations, which are arranged on strips on the box wall.

[0186] Figure 28 The mean square error of the temperature output by the PIML model is shown as a function of the training period.

[0187] Figure 29 This is a flowchart of the method.

[0188] Figure 30 It is a schematic representation of another system that includes the processing system.

[0189] Figure 31 It is a schematic representation of another system that includes the processing system.

[0190] Figure 32 It is a schematic representation of another system that includes the processing system.

[0191] Figure 33 It is a schematic representation of another system that includes the processing system.

[0192] Figure 34 It is a schematic representation of another system that includes the processing system.

[0193] Figure 35 It is a schematic representation of another system that includes the processing system. Detailed Implementation

[0194] Embodiments of the invention will be described with reference to the accompanying drawings. In the drawings, similar or identical reference numerals denote elements having similar or identical configurations and / or functions.

[0195] The embodiments relate to processing systems and methods capable of operating to determine one or more parameters of an asset of a power system in a spatially resolved manner, and optionally in a spatiotemporally resolved manner.

[0196] These processing systems and methods are operable to perform physical information data-driven processing techniques (i.e., processing techniques with embedded knowledge of physical relationships) to determine parameters. Data-driven processing techniques may include Physical Information Machine Learning (PIML) models, such as Physical Information Neural Networks (PINNs). PIML models may have outputs operable to provide parameters. PIML models may have inputs operable to receive input data, including the location and optionally the time at which the parameters are to be determined, measurements of locations where the measurements are considered reliable (e.g., thermal measurements), and / or at least one operating parameter (e.g., load power and / or load current).

[0197] Although embodiments will be described in detail primarily in relation to power systems, including transmission and / or distribution systems, the embodiments are not limited thereto.

[0198] The processing methods and systems disclosed herein are operable for processing measurement results obtained on assets with enclosures (such as transformers or shunt reactors including enclosures, which may be filled with insulating fluids (e.g., insulating oil)). Physical information data-driven processing techniques are operable for processing measurement results obtained on at least one or more sidewalls of the enclosure. These techniques are operable for determining temperature and / or heat in a spatially resolved, optionally spatiotemporally resolved manner. The processing system is then operable for generating an output based on the determined spatially resolved temperature and / or heat.

[0199] As used herein, the term "electric system" encompasses generation systems, transmission systems, and / or distribution systems, such as a power grid or a portion thereof.

[0200] As used herein, the term "power grid" encompasses both transmission and / or distribution networks. A power grid may include both transmission and distribution networks.

[0201] A physical information data-driven processing technique capable of determining at least one parameter can be operated to provide at least one parameter as output in a spatially resolved manner. The physical information data-driven processing technique can be operated to provide a first parameter and a second parameter as output in a spatially resolved manner (optionally in a spatiotemporally resolved manner), wherein the first parameter and the second parameter are correlated with each other through physical relationships (such as diffusion equations, e.g., thermal diffusion equations) embedded in the physical information data-driven processing technique. The physical information data-driven processing technique may have an input capable of receiving input data indicating one or more locations (e.g., on the wall of an asset), a measurement result of the first parameter (if available for the corresponding location), and at least one operational parameter affecting the second parameter.

[0202] As used herein, the “operating parameters” that can be received at the input of physical information data-driven processing techniques (e.g., at the input of a PIML model) can be any operating parameters that affect parameters determined using data-driven processing techniques. For illustration, load power or load current are examples of operating parameters that are useful when applying the techniques disclosed herein to temperature-based asset monitoring applications capable of identifying localized losses and triggering appropriate actions in response to those losses.

[0203] The techniques disclosed herein are operable to perform preprocessing to transform data received by a processing system into input data suitable for input into physical information data-driven processing techniques. Preprocessing includes determining which measurement results are reliable, and the input data is generated based on the reliable measurements and independently of those not determined to be reliable.

[0204] As used herein, a measurement result is determined to be “reliable” when it reflects the physical relationships embedded in the physical information data-driven processing technique, or a measurement location is determined to provide “reliable” measurement results. In other words, measurement results or measurement locations are considered “reliable” when they do not exhibit behavior that contradicts the physical relationships embedded in the physical information data-driven processing technique. For illustration, the physical information data-driven processing technique can operate to determine several parameters that are related to each other through diffusion equations (e.g., thermal diffusion equations). In this case, a measurement result of the first parameter (such as temperature) can be identified as “unreliable” when the first derivative exhibits discontinuities (in which case, the second derivative in the diffusion equation would be poorly defined or divergent, which is anomalous). Such a measurement result being determined as “unreliable” does not indicate an incorrect temperature measurement, but rather that these measurements cannot be relied upon to generate input data for the physical information data-driven processing technique.

[0205] Examples of such measurement results that are unsuitable for generating input data for physical information data-driven processing technologies may include measurements taken at locations where electrical and / or mechanical components within the structural components and / or walls of the enclosure act as radiators and / or thermal radiation attenuators, making thermal measurements unreflective of the thermodynamics (such as thermal diffusion) within the asset.

[0206] Preprocessing may include transforming the thermal camera output to map pixel coordinates to world coordinates within the asset coordinate frame.

[0207] Preprocessing may include using separate meshes for structural components (such as supports) of the asset's box in cases where measurement results are unreliable. Alternatively or additionally, preprocessing may include edge detection, for example, edge detection of sections where measurement results are unreliable or redundant.

[0208] Preprocessing may include removing pixels near the stripe boundaries and performing interpolation or extrapolation to ensure the continuity of heat diffusion.

[0209] Physical information data-driven processing techniques can be trained using measurements obtained on the same asset to which the physical information data-driven processing techniques are applied during inference and / or on another asset with a similar construction (e.g., when the asset is a transformer with a box wall, the other asset is another transformer with a box wall).

[0210] Physical information data-driven processing techniques can be trained using causal training, which penalizes high training loss at any past time step. Alternatively or additionally, temporal gradient regularization can be used in training to penalize high loss and low derivatives of the loss function (e.g., low gradients along the time axis).

[0211] Physical information data-driven processing technology can be used to verify the results of a first parameter measurement (such as spatially resolved temperature measurement) and a second parameter simulation result (such as heat). For the first parameter, such a measurement result is available; for the second parameter, such a measurement result is not available, but is still determined by the physical information data-driven processing technology. To perform this verification, the computational system performing the verification can operate to interface with a simulation system (such as a Python-MATLAB-COMSOL-based simulation system).

[0212] Various effects and advantages are associated with these processing methods and systems. For illustration, these processing systems and methods can be operated to provide parameter values ​​with enhanced robustness compared to conventional techniques.

[0213] These processing systems and methods can be operated to utilize PIML. PIML is an emerging field in machine learning (ML) where the underlying physics of a system is estimated correlated with observed data, making the trained model less dependent on data because, compared to a purely data-driven process, the trained model learns to make predictions directly from the dynamics. Neural networks (NNs) (such as deep or shallow NNs) that can be trained to estimate dynamics as well as observed variables are called Physical Information Neural Networks (PINNs), and they can be used to determine parameters by learning from fully known dynamics (forward problem) or estimating unknown or uncertain components of partially known dynamics (backward problem). Because PINNs are able to estimate dynamics using limited observations, they can be used to estimate the dynamics in complex devices where the dynamics are partially known, making simulation difficult or impossible due to a lack of knowledge of the physical model, or where data measurement is costly.

[0214] These processing systems and methods address the challenge that, in some real-world applications, the availability of measurable data may be limited and / or the data quality may be insufficient. One example is the impact of structural components of an asset enclosure (such as enclosure reinforcements) on temperature measurements. These processing systems and methods provide techniques for estimating thermal diffusion in complex electromagnetic devices (such as power transformers or shunt reactors) using PIML models (e.g., PINN) in the presence of limited and / or unreliable measurement results. To this end, these processing systems and methods are capable of manipulating inputs used to perform preprocessing to generate the PIML model.

[0215] Figure 1 A schematic representation of system 10 is shown. System 10 includes power system assets 20, measuring instruments 41, 42, and a processing system 50 capable of operating to process the measurement results. The processing system 50 is capable of operating to perform physical information data-driven processing techniques to determine one or more parameters, such as temperature and heat, in a spatially resolved manner, and optionally in a spatiotemporally resolved manner.

[0216] Asset 20 has a box 30. Box 30 has multiple side walls 31, 32, a lid 35, and a bottom (which is located in...). Figure 1(Not visible in the view). Asset 20 may have structural components, such as box reinforcements 36 provided at least along various portions of the box wall. These structural components may reduce the significance of thermal measurements at the location of the structural components on the box wall and in adjacent areas. Asset 20 may have accessory components, such as sleeve 21 and expansion tank 22. Box 20 may be filled with an insulating liquid, such as insulating oil. Expansion tank 22 is operable to receive oil from box 20, which is fluidly connected to it. Thus, even if the density of the oil changes due to thermal effects, a sufficient amount of oil can be ensured to be present in box 20.

[0217] The measuring instrument may include one or more sensors operable to sense temperature at multiple locations on the walls of the enclosure 30. The measuring instrument may be operable to sense, or indicate the amount of temperature, at least two side walls 31, 32 and optionally the floor cover 35 in a spatially resolved manner. The measuring instrument may include at least one thermal camera 41, 42. The measuring instrument may include one or more thermal cameras 41, 42 operable to perform measurements on at least two side walls 31, 32 of the enclosure 30. The measuring instrument may additionally or alternatively include other sensors, such as optical or fiber optic sensors, thermocouples, or other temperature sensing devices operable to measure temperature.

[0218] The processing system 50 includes at least one interface 51. The at least one interface 51 is operable to receive data based on measurement results captured by a measuring instrument. The at least one interface 51 can be connected to the measuring instrument via at least one communication link 43, 44. The at least one communication link 43, 44 can be a unidirectional or bidirectional communication link. In the latter case, the processing system 50 can be operable to control the measuring instrument to perform measurements.

[0219] Typically, at least one interface 51 of the processing system 50 is operable to receive data 58, including or otherwise based on measurement results. The data 58 may also optionally include at least one operating parameter of the asset 20, such as load current and / or load power. At least one interface 51 of the processing system 50 is operable to provide an output 59. The output 59 may be based at least on one or more determined parameters.

[0220] Processing system 50 typically includes one or more processing circuits 60 and at least one interface 51. Processing system 50 may optionally include a human-machine interface (HMI) 52. Processing system 60 may include or be communicatively coupled to storage system 53. Storage system 53 may store parameters, and optionally hyperparameters of a physical information data-driven processing technique executed by processing system 50. Parameters and optionally hyperparameters may define a PIML model, such as PINN, which is operable to determine at least one parameter in a spatially resolved manner. At least one processing circuit 60 may be operable to retrieve parameters and optionally hyperparameters from storage system 53. Parameters and optionally hyperparameters stored in storage system 53 may be determined during training. Training may be implemented in a manner that encodes physical relationships in a trained data-driven processing technique. Thus, the processing system 50 is capable of operating to apply physical information data-driven processing techniques in a manner that provides physically plausible solutions for a first parameter (e.g., temperature) and a second parameter (e.g., heat) (such as spatiotemporal solutions of temperature and heat considering the heat diffusion equation). The input data for the physical information data-driven processing techniques is generated using preprocessing based on measurements of the first parameter (and does not require measurements of the second parameter), which considers whether the measurements of the first parameter can represent the physical behavior embedded in the physical information data-driven processing techniques. For example, measurement locations where the measurements cannot represent physical relationships due to divergent and / or poorly defined first spatial derivatives can be discarded to generate input data.

[0221] At least one processing circuit 60 is operable to perform measurement processing 61. Measurement processing 61 may include preprocessing 70. Preprocessing 70 may be operable to transform measurement results obtained by measuring instruments 41, 42 into a desired coordinate system (such as a coordinate system fixed to the sidewalls 31, 32 of housing 30). Preprocessing 70 includes input data generation 71, which is operable to generate input data for physical information data-driven processing techniques. Input data generation 71 may include identifying which measurement results are reliable, in this sense representing physical relationships embedded in physical information data-driven processing techniques. Preprocessing 70 may crop measurement regions where low reliability is expected, such as measurement results in areas of structural component 36, which can be identified because they are line-bounded by discontinuities exhibited by the spatial derivative of temperature (in this case, the second derivative is poorly defined or divergent).

[0222] At least one processing circuit 60 is operable to perform a determination 62 of at least one parameter. At least one processing circuit 60 is operable to cause the determination 62 to determine both heat and temperature in a spatially resolved manner, and optionally in a spatiotemporally resolved manner.

[0223] At least one processing circuit 60 is operable to perform interface control 69. Interface control 69 is operable to control at least one interface 51 and / or HMI 52 to provide output based on at least one determined parameter.

[0224] To perform the above operations, at least one processing circuit 60 may include any one or any combination of an integrated circuit, an integrated semiconductor circuit, a processor, a controller, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a circuit including qubits and / or quantum gates, or a combination thereof, but is not limited thereto.

[0225] The output can be generated in various ways using at least one determined parameter (which can be spatially resolved on at least a portion of the side walls 31, 32 of the enclosure 30). For illustration, this at least one parameter can be used to identify which spatial partitions on the enclosure 30 are subject to localized heat and / or temperature indicators determined using physical information data-driven processing techniques. Alternatively or additionally, this at least one parameter can be used to determine whether the identified localized increase in heat and / or temperature warrants automatic control action and / or the output of warnings, alerts, or alarms.

[0226] Therefore, at least one interface control 69 can operate to control at least one interface 51 and / or HMI 52 to provide outputs dependent on at least one determined parameter.

[0227] Output 59 may include control commands or other control data that triggers actions affecting major systems of the power system, including asset 20. Processing system 50 may be operable to directly generate and provide control commands. Alternatively or additionally, processing system 50 may be connected to the control system 49 of the power system via a communication interface. Processing system 50 may be operable to generate output 59 such that the control system 49 causes the control system 49 to perform actions affecting major system equipment of the power system.

[0228] Figure 2This is a schematic block diagram representation of measurement processing 61. Preprocessing 70 receives input 64 based at least on measurement results obtained using measuring instruments. Input 64 may include temperature measurements at multiple locations on the walls of chamber 30. Preprocessing 70 may perform one or more operations to generate PIML model input 65 based on input 64. Preprocessing 70 may include identification 72 of reliable and / or unreliable measurement results. Identification 72 may include one or more of the segmentation to identify structural components of the chamber wall that may render the thermal measurement results unreliable. Identification 72 may be performed based on spatially resolved temperature measurement results. For illustration, a discontinuity in the first spatial derivative of the spatially resolved temperature measurement results may indicate that at the location where the discontinuity is detected, there may be structural components of the chamber wall and / or components within the chamber wall that have the effect of not indicating heat diffusion in the measurement results (e.g., due to the presence of radiators and / or thermal radiation attenuators). Preprocessing 70 includes generating 73 input data based on the identification. The generation of input data 73 may include removing measurements deemed unreliable (e.g., due to the presence of structural components of box 30) and / or performing operations such as extrapolation, interpolation, or regression to replace measurements that may be identified as unreliable with alternative data that can be updated from neighboring regions where thermal measurements are more reliable.

[0229] The determination of at least one parameter 62 includes processing the input data 65 using a physical information data-driven processing technique 63. The physical information data-driven processing technique may include a PIML model, such as PINN. As will be described in more detail below, the physical information data-driven processing technique 63 operates in such a spatially resolved manner that it determines at least one parameter (such as temperature and / or heat).

[0230] Determined 62 is operable to provide at least one parameter 66, for example, as the output of a PIML model. The at least one parameter 66 may be provided for each of a plurality of locations on at least two sidewalls 31, 32 of the box 30, and optionally for a plurality of time points.

[0231] The processing system 50 can be operated to determine at least one parameter using a PIML model.

[0232] Figure 3 This is a schematic representation of a measurement process 61 that can be performed by at least one processing circuit 60 of the processing system 50. The measurement process 61 includes preprocessing 70. Preprocessing 70 may include cropping and / or coordinate transformation 74 of the measurement data. For illustration, when using a thermal camera, preprocessing may include cropping 74 to remove pixels or voxels that do not represent any part of the housing 30 and / or transforming the thermal image into a coordinate frame attached to one or more sidewalls 31, 32 of the housing 30.

[0233] Preprocessing 70 may include segmentation 75. Segmentation 75 may be performed based on thermal measurement results. Segmentation 75 may include identifying pixels or voxels arranged at structural components of the housing 30 based on the measurement results, which reduce the reliability of the thermal measurement results, in the sense that the thermal measurement results do not indicate diffusion processes (e.g., thermal diffusion) embedded in the physical information data-driven processing technique 63.

[0234] Preprocessing 70 may include interpolation, extrapolation, and / or other filtering operations 76. The interpolation, extrapolation, and / or other filtering operations are preferably local, in that sense only considering pixels in the local environment of the pixels replaced by the interpolated, extrapolated, or other filtered values. Thus, the local correlation between measurements and positions on the bin 30 remains available during subsequent processing using the physical information data-driven processing technique 63.

[0235] The output 67 of the physical information data-driven processing technique 63 can be processed to generate an output 66, wherein the output 66 is operable to trigger an action (such as an output or automatic control action). For illustration, the output 67 of the physical information data-driven processing technique 63 can provide temperature and / or heat in a spatially resolved manner, optionally in a spatiotemporally resolved manner. Determining at least one parameter can include determining 64 the location and / or magnitude of localized damage to the asset, as indicated by heat. Output 66 can rely on this identification of localized damage.

[0236] Figure 4 This is a flowchart of method 80. Method 80 can be executed automatically by the processing system 50.

[0237] At process block 81, the measurement results are preprocessed. Preprocessing the measurement results may include converting the measurement results into a format suitable for being provided as input data to the PIML model. Preprocessing includes identifying reliable measurement results and using these reliable measurement results to generate input data for processing by physical information data-driven processing techniques. Identifying reliable measurement results may include identifying measurement locations where the measurement results have partial spatial derivatives and / or partial temporal derivatives that are not inconsistent with physical relationships embedded in the physical information data-driven processing techniques. Identifying reliable measurement results may include identifying measurement locations where the measurement results have first-order partial spatial derivatives and / or temporal derivatives that do not exhibit discontinuities. Preprocessing may include segmenting and / or cropping the thermal image. Preprocessing may include identifying regions where thermal measurement results are unreliable due to the presence of structural components, such as those in box 30, that provide heat sinks and / or thermal shielding. Preprocessing may include interpolation or extrapolation to facilitate subsequent processing of the PIML model.

[0238] At process block 82, physical information data-driven processing techniques are used to determine at least one parameter. For example, at multiple locations on the side walls 31, 32 of box 30, at least one parameter can be determined in a spatially resolved manner, or optionally in a spatiotemporally resolved manner.

[0239] At process block 83, the processing system causes at least one action to be performed based on at least one determined parameter. Performing at least one action may include triggering a warning, alert, or alarm. Alternatively or additionally, performing at least one action may include generating control commands or triggering control data that changes operating conditions (such as changes affecting major system equipment of the power system).

[0240] Further reference Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 and Figure 11 The diagram illustrates the operation of the processing system.

[0241] Figure 5 Measurement results 91 are shown, including thermal images acquired on the sidewalls 31 of the asset's box 30. The arrangement and orientation of the sensors (in this case, thermal cameras) relative to the box 30 can cause distortions and may also cause some pixels to represent the background instead of the sidewalls 31. Preprocessing performed by the processing system at least partially compensates for this effect caused by the potentially varying sensor arrangement and orientation. For this purpose, preprocessing may include cropping and / or coordinate transformation. Measurement results 91 also include unreliable measurements in regions 101, 102, where these measurements are unreliable in the sense that the measurements exhibit behavior inconsistent with the physical relationships embedded in the data-driven processing technique (such as discontinuities in spatial and / or temporal derivatives). Preprocessing identifies such regions and generates input data for the physical information data-driven processing technique in such a way that the input data does not depend on measurements inconsistent with the physical relationships embedded in the data-driven processing technique.

[0242] Figure 6 A cropped thermal image 92 is shown, in which a coordinate transformation is performed to map the pixels of the thermal image to associated coordinates on the transformer wall. Figure 6 As shown, there are significant temperature drops in certain areas, caused by structural components of the enclosure walls rather than the internal thermodynamics of the asset. To reduce or eliminate the potential impact of this effect, preprocessing includes removing or reducing the segmentation of such areas where thermal measurements are inconsistent with the physical relationships embedded in the data-driven processing techniques (due to, for example, the presence of structural components at the enclosure walls and / or components within the enclosure that act as radiators and / or attenuate thermal radiation).

[0243] Figure 7 The preprocessed measurement results 93 are shown after removing regions inconsistent with the physical relationships embedded in the data-driven processing technique.

[0244] The results of the preprocessing can be used in PIML models or other physical information data-driven processing techniques that can operate to predict at least one or more parameters (such as temperature and heat) in a spatially resolved manner.

[0245] Various measurement tools can be used to collect data from assets such as power transformers or shunt reactors. Figure 5 Temperature data is required, and the applicability of these measuring tools may vary depending on the availability of the component. Measurement results may include heat distribution on the walls of an electrical installation captured using a thermal camera and / or thermocouples. The acquired thermal images can be transformed into a two-dimensional view to account for translational and rotational shift errors during data capture. Figure 6 Then, the two-dimensional corrected view can be further divided into strips by removing portions that do not reflect the effect of thermal diffusion on the surface. Figure 7 When the target area of ​​asset 20 cannot be captured by a single camera due to insufficient distance from the object to the camera or due to obstacles, individual measurement results from different viewpoints or angles can be performed and processed. The individual measurement results are projected onto the corresponding positions in the two-dimensional view so that the entire surface of the measured target is covered.

[0246] Depending on the scenario in which temperature measurements are collected, such as factory testing or on-site operational monitoring, thermal images may capture measurements that do not reflect the thermal diffusion processes within an asset. To address this, such pixels are preprocessed to identify them and can be discarded when generating input data for physical information data-driven processing techniques.

[0247] Figure 8 , Figure 9 and Figure 10 Alternative illustrations are provided for pixels that do not represent thermal diffusion embedded in physical information data-driven processing techniques. This identification can be performed on the extracted strips using edge detection techniques. Figure 8 It shows Figure 7 The cropped segmentation data 93 is one of the amplified portions of the stripe 94, and Figure 9 and Figure 10 It shows Figure 8 Details 97 and 98 in areas 95 and 96. Figure 8 , Figure 9 and Figure 10 This shows the use of pixel gradients. - An example of edge detection using the -norm. Interpolation and / or extrapolation can be performed.

[0248] Figure 11 It shows Figure 8 The interpolated data for part 94, where interpolation is performed using cubic splines after removing pixels detected as edges.

[0249] Therefore, pixels that do not reflect thermal diffusion can be identified and removed from consideration. The processing system 50 can be operated to control the HMI 52 to enable user input specifying which portions to discard when generating input data for the physical information data-driven processing technique. For rectangular portions, the processing system 50 can be operated to control the HMI 52 to enable user input specifying corner positions. For circles or ellipses, the processing system 50 can be operated to control the HMI 52 to enable user input specifying the center position and at least one radius.

[0250] Figure 12 This is a flowchart of method 110. Method 110 can be executed automatically by processing system 50. Method 110 can be executed to perform preprocessing at process block 81 of method 80. Method 110 can be executed to generate input data for a physical information data-driven processing technique capable of providing temperature and heat in a spatially resolved manner (optionally in a spatiotemporally resolved manner) and trained with a loss function that depends on the residuals of a linear PDE representing heat diffusion.

[0251] At process block 111, a coordinate transformation is performed on the received measurement results. The coordinate transformation may include rotation and / or translation. The coordinate transformation may depend on the position and angular orientation of the measuring instruments 41, 42 relative to asset 20 (e.g., relative to box 30).

[0252] At process block 112, one or more regions where the measurement results do not indicate thermal diffusion are identified. Identification of these regions can be performed based on the spatial gradient of the measurement results. Identification can also be performed based on edge detection, which can be based on, for example, the L1 norm.

[0253] At process block 113, when input data is generated, the processing system discards or replaces measurement results obtained in one or more regions where the temperature measurement results do not indicate thermal diffusion.

[0254] Discarding may include inputting spatial coordinates into physically-informed data-driven processing techniques (e.g., inputting them into a PIML model). - If the spatial coordinates are within the region indicated by the temperature measurement results for thermal diffusion, then use temperature markings, and - If the spatial coordinates are in an area where the temperature measurement results do not indicate thermal diffusion, then temperature marking is not required.

[0255] The replacement may include inputting spatial coordinates into a physically-informed data-driven processing technique (e.g., inputting them into a PIML model). - If the spatial coordinates lie within the region indicated by the temperature measurement results for thermal diffusion, then the temperature value obtained through measurement is marked, and - If the spatial coordinates are in an area where the temperature measurement results do not indicate thermal diffusion, then the temperature value is marked by interpolation or extrapolation from other temperature values.

[0256] Temperature and / or heat can be determined in a spatially resolved manner because the processing system 50 is capable of operating to input various spatial coordinates (e.g., spatial coordinates on one or more of the side walls and / or lids of the box 30) and at least one operating parameter (such as load current and / or load power) into the data-driven processing technology, where applicable, optionally having temperature labels.

[0257] Temperature and / or heat can be determined in a spatiotemporally resolved manner because the processing system 50 is operable to input at least one operating parameter (such as load current and / or load power) to a data-driven processing technique, which may optionally have temperature labels (e.g., spatial coordinates on one or more of the side walls and / or lids of the enclosure 30) and various times and corresponding times. Thus, temperature and heat can be determined in a spatiotemporally resolved manner using the data-driven processing technique.

[0258] Figure 13 This is a flowchart of method 115. Method 115 can be executed automatically by processing system 50. Method 115 can be executed to perform preprocessing at process block 81 of method 80. Method 115 can be executed to generate input data for a physical information data-driven processing technique capable of providing temperature and heat in a spatially resolved manner (optionally in a spatiotemporally resolved manner) and trained with a loss function that depends on the residuals of a linear PDE representing heat diffusion.

[0259] Process blocks 111, 112, and 113 can be referenced as previously. Figure 12 Implement it as explained.

[0260] At process block 116, method 115 may further include edge detection for determining structures where temperature measurements do not indicate (e.g., along the chamber wall) thermal diffusion.

[0261] At process block 117, method 115 may further include interpolation, extrapolation, and / or application of a local filtering kernel at the location determined by edge detection at process block 116.

[0262] Figure 14 and Figure 15An implementation of a physical information data-driven processing technique implemented in a PIML mode (such as PINN) is shown.

[0263] Figure 14 This is a schematic representation of a PIML model 120. The PIML model 120 has inputs 122 and 126. The PIML model 120 includes a PIML model 120 portion 121 operable for predicting a first parameter (e.g., temperature). The PIML model 120 may include an additional PIML model 120 portion 125 operable for predicting a second parameter (e.g., a seed) related to the first parameter by means of a physical relationship. The PIML model 120 may have an output 123 operable for providing the first parameter. The PIML model 120 may have an additional output 127 operable for providing the second parameter.

[0264] PIML model 120 is trained by embedding physical relations. These physical relations depend on a first parameter and a second parameter. For the thermal diffusion phenomenon, the physical relations can be partial differential equations (PDEs) that correlate the first partial time derivative of temperature, the second partial spatial derivative of temperature (such as a 2D or 3D Laplace operator), and the rate of change of heat. This physical relation embedded in PIML model 120 is schematically illustrated by operator 128 and summation / subtraction nodes 129. The output of summation node 129 provides the residuals of the PDE. During training, PIML model 120 is trained using a loss function that includes terms dependent on the PDE residuals. Thus, the first and second PIML model parts 120 121, 125 are trained to provide physically meaningful solutions that conform to the thermal diffusion equation.

[0265] What will be understood is that during inference, it is not necessary (but still possible) to determine the residuals of the PDE. In other words, when using PIML model 120 during inference, operations 128 and 129 can be omitted.

[0266] Input node 122 can be operated to receive spatial coordinates (such as a pair of coordinates for a location on the sidewalls 31, 32 of a specified asset) and time. Where applicable and deemed reliable, the temperature and time at this location can be used as labels for the input, or can otherwise be provided at input node 122. It should be understood that although temperature measurements are obtained using measuring instruments, the PIML model 120 provides additional useful information because, firstly, it provides an overall temperature and heat determination consistent with heat diffusion, and secondly, it can also provide temperature and / or heat in areas where the measurements do not indicate heat diffusion.

[0267] Input node 126 can be operated to receive spatial coordinates and operating conditions affecting temperature and / or heat. For illustration, input node 127 can be operated to receive spatial coordinates (such as a pair of coordinates specifying the location on the sidewalls 31, 32 of the transformer) and load power and / or load current.

[0268] Figure 15 This is a schematic representation of a PIML model 130. The PIML model 130 includes a first plurality of neural network (NN) layers 131 having a first input layer 132 and a first output layer 133. The PIML model 130 includes a second plurality of NN layers 134 having a second input layer 135 coupled to the first output layer 133. The second plurality of NN layers 134 has an output layer 136 capable of operating to provide a first parameter (e.g., temperature). The PIML model 130 includes a third plurality of NN layers 137 having a third input layer 138 coupled to the first output layer 133. The third plurality of NN layers 137 has an output layer 139 capable of operating to provide a second parameter (e.g., heat). The PIML model 130 is trained such that the outputs at output nodes 136, 139 conform to (e.g., by using a PIML model trained with a loss function that depends on the residuals of a heat diffusion PDE) a physical relationship (such as heat diffusion) embedded in at least one PIML model.

[0269] For reference Figure 15 The explained configuration offers various technical effects. Training two separate architectures for predicting temperature and heat requires more memory and more training time, which can be significant depending on the use case, especially with a larger required size for each model. The PIML model architecture with a common NN layer of 131 has the effect of sharing some PIML model parameters. The degree of sharing is a hyperparameter that can be based on, for example, accuracy and convergence time based on the use case. Figure 15 In the diagram, the number of layers in each of sections 131, 134, and 137 are hyperparameters that can be optimized during training. Alternatively or additionally, the number of neurons per layer can be different.

[0270] To further enhance the accuracy obtained from the PIML model, the PIML model can have an output layer that mandates positive values ​​for the temperature and heat supplied at the output. Since the PIML model output in the inverse problem of heat diffusion is temperature (typically in Kelvin) and load loss, both of which are non-negative, the model's output can be forced to be non-negative by introducing absolute values ​​as the final layer of the PIML model. This reduces estimation errors.

[0271] Figure 16This illustrates the effect of the deviation between the heat (load loss) output by the PIML model and real data obtained through simulation (where real data is not used during training). Figure 16 Curves 141 and 142 in the diagram are intercepted along the line of maximum deviation on the box wall 30. Curve 141 represents the deviation of the PIML model, which has an output layer that mandates positive temperature and heat properties. Curve 142 represents the deviation of the PIML model where positive properties are not mandated during training. Figure 16 It can be seen that accuracy is improved by using a PIML model that forces the output (temperature and heat in Kelvin) to be non-negative. It is worth emphasizing again that... Figure 16 In the simulation, the actual data value of heat loss is... Used only for validation, not during training.

[0272] Figure 17 This is a flowchart of method 150. Method 150 can be executed automatically by processing system 50.

[0273] At process frame 151, the processing system obtains the spatiotemporally resolved thermal data of the asset's box walls.

[0274] At process block 152, the processing system acquires at least one operating condition that affects heat generation in the asset. Load current and / or load power are examples of such operating conditions.

[0275] At process block 153, the processing system performs preprocessing of the thermal data. Preprocessing may include identifying one or more areas where the thermal data is unreliable, and generating input data for physical information data-driven processing techniques based on determinations.

[0276] At process block 154, the processing system processes the input data to determine the temperature and heat in a spatially resolved manner, and optionally in a spatiotemporally resolved manner. The processing system can use, as in reference... Figure 14 or Figure 15 The PIML model is used to perform the processing.

[0277] At process block 155, the processing system causes at least one action to be performed based on determined parameters (temperature and / or heat).

[0278] The preprocessing disclosed in this paper can be performed not only during inference but also during training. This enhances prediction accuracy. Other aspects that can be used in conjunction with training have been explained, such as using a PIML model with output layers that have parameters that are known to be nonnegative, and / or determining the number of layers and / or neurons per layer as hyperparameters during training. Further aspects related to training and / or validation will be described in more detail below. Training can be performed by processing system 50 or by another computing system that can be connected to processing system 50 via a communication interface or otherwise operable to provide PIML model data to processing system 50.

[0279] Figure 18 This is a flowchart of method 160, which provides a PIML model for use by processing system 50. Method 160 can be executed by processing system 50 or by another computing system.

[0280] At process box 161, obtain training data. Training data may include measurements of at least a first parameter (such as temperature) on an asset that is the same as, or similar in construction to, the asset on which the PIML model will be applied.

[0281] At process block 162, preprocessing of the training data is performed. Preprocessing of the training data may include the methods described in this paper. Figures 1 to 17 Any operations that interpret this. Preprocessing of training data may specifically include determining, based on measurements of the first parameter, which regions (such as regions) do not indicate physical processes (such as thermal diffusion) embedded in the PIML model, and replacing or discarding such measurements that are inconsistent with that physical process.

[0282] At process block 163, the PIML model is trained using preprocessed training data. Training the PIML model may include training with a loss function that depends on the estimation error and the PDE residuals. Training may be performed as history-aware training, which depends on the derivatives of the loss function at previous training times or preferably several previous training times.

[0283] Figure 19 It is a schematic representation of a system including a computing system 170, which is operable to provide a PIML model for use by a processing system 50.

[0284] The computing system 170 includes at least one interface 171. The at least one interface 171 is operable for receiving training data based on measurement results captured by a measuring instrument. The at least one interface 171 can communicate with the measuring instrument and / or with a data repository storing the training data.

[0285] Typically, at least one interface 171 of the computing system 170 is operable to receive input 178, which includes or is otherwise based on measurement results. Input 178 may also optionally include at least one operating parameter, such as load current and / or load power. At least one interface 171 of the computing system 170 is operable to provide output 179. Output 179 may be based at least on the results of PIML model training. Output 179 may include PIML model parameters and / or hyperparameters for use by the processing system 50.

[0286] The computing system 170 typically includes one or more circuits 173 and at least one interface 171. The computing system 170 may include a storage system 172, or may be communicatively coupled to a storage system 172. The storage system 172 may be used to store training data and / or parameters, and / or data that can be used to perform historically-aware training (such as loss function values ​​and / or the time derivative of the loss function determined in previous time instances of training). The storage system 172 may be operable to store parameters and / or hyperparameters of a PIML model, such as those determined by the computing system 170. The parameters, and optionally the hyperparameters, may define a PIML model, such as PINN, which is operable to determine at least one parameter (such as temperature and heat) in a spatially resolved manner.

[0287] At least one circuit 173 is operable to perform PIML model training and validation 174 based on input 178. PIML model training and validation 174 includes preprocessing 175. This can be achieved using methods such as those described in the references. Figures 1 to 18 The described technique is used to implement preprocessing 175. Preprocessing 175 may include identifying which measurement results included in input 178 are reliable, in the sense that they can represent physical relationships embedded in physical information data-driven processing techniques. Preprocessing 175 may prune measurement regions where low reliability is expected, such as measurement results in the region of structural component 36, which can be identified because they are line-bounded by discontinuities exhibited by the spatial derivative of temperature (in this case, the second derivative is poorly defined or divergent).

[0288] At least one circuit 173 is operable to perform PIML model training 176. PIML model training 176 can be performed using a loss function that depends on the estimation error (e.g., mean squared error) and the PDE residuals. The loss function can be set in such a way that the loss function at a certain time depends on the loss function of one or preferably several previous time instances, in particular the time derivative of the loss function.

[0289] At least one circuit 173 is operable to perform PIML model verification 177. The computing system 170 is operable to establish a communication interface with at least one simulation system 169 to perform the verification. For illustration, when measurement results of real data for one of the parameters output by the PIML model are unavailable, at least one circuit 173 is operable to provide the spatially varying temperature determined by the PIML model to at least one simulation system 169 via at least one interface 171, and to receive simulation results of real heat data from at least one interface 171. The simulation results of the real data can be used for verification.

[0290] To perform the above operations, at least one circuit 173 may include any one or any combination of an integrated circuit, an integrated semiconductor circuit, a processor, a controller, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a circuit including qubits and / or quantum gates, or a combination thereof, but is not limited thereto.

[0291] Training can be performed to jointly (or collaboratively) train all parameters of the PIML model. However, it has been found that training can be performed efficiently, and accuracy can be improved by training the parts related to different parameters separately before jointly training the parts related to determining temperature and the parts related to determining heat.

[0292] Figure 20 This is a flowchart of method 180 for providing a PIML model. Method 180 can be executed to implement step 163 of method 160. Method 180 can be executed automatically by computing system 170.

[0293] At process block 181, the first NN part of the PIML model is trained to determine the first parameter (e.g., temperature), while keeping the parameters of the second NN part of the PIML model unchanged, which determines the second parameter (e.g., heat).

[0294] At process block 182, the second NN part of the PIML model is trained to determine the second parameter (e.g., heat), while keeping the parameters of the first NN part of the PIML model unchanged.

[0295] At process frame 183, the first NN part and the second NN part are trained together.

[0296] For illustrative purposes, see reference Figure 14 Training may include training the first NN part 121 while keeping the parameters of the second NN part 125 fixed, training the second NN part 125 while keeping the parameters of the first NN part 121 fixed, and then co-training the first NN part 121 and the second NN part 125.

[0297] For further explanation, see reference Figure 15 Training may include training NN parts 131 and 134 while keeping the parameters of NN part 137 fixed, training NN parts 131 and 137 while keeping the parameters of NN part 134 fixed, and then co-training NN parts 131, 134 and 137.

[0298] It has been found that training speed is significantly improved by training each part separately before co-training. To illustrate, the temperature neural network (NN) can be trained first without updating the parameters in the heat neural network (NN). The heat NN is then trained for a period while the parameters in the temperature NN are no longer updated. Once both NNs are trained to reduce transient training error, they are trained together for the remainder of the training period. Training with NNs pre-trained in this way shows enhanced stability at the beginning of the training period, leading to faster convergence.

[0299] Figure 21 This is a flowchart of method 165, which provides a PIML model for use by processing system 50. Method 165 can be executed by processing system 50 or by computing system 170.

[0300] At process box 161, obtain training data. Training data may include measurements of at least a first parameter (such as temperature) on an asset that is the same as, or similar in construction to, the asset on which the PIML model will be applied.

[0301] At process block 162, preprocessing of the training data is performed. Preprocessing of the training data may include the methods described in this paper. Figures 1 to 17 Any operations that interpret this. Preprocessing of training data may specifically include determining, based on measurements of the first parameter, which regions (such as regions) do not indicate physical processes (such as thermal diffusion) embedded in the PIML model, and replacing or discarding such measurements that are inconsistent with that physical process.

[0302] At process block 165, the PIML model is trained using preprocessed training data, where training includes history-aware training. In history-aware training, the previous loss from previous training times or several previous training times is considered.

[0303] History-aware training provides various technical effects when used in conjunction with the PIML model disclosed in this paper. Since estimation errors can propagate through the underlying PDE in a dynamical system, the prediction accuracy of the dynamics can be affected by past estimation errors. If the estimation errors in one or more earlier epochs of training are quite high, it will also affect the prediction accuracy in later epochs, and consequently, the convergence speed.

[0304] One way to force a lower impact from past errors is to penalize high past errors.

[0305] In this first implementation of history-aware training, causal training following the approach of S. Wang et al., “Respecting causality is all you need for training physics-informed neural networks,” March 14, 2022, arXiv:2203.07404, is applied to penalize all past moments with large losses by introducing an exponential weight to each time loss, which is a function of all past losses. In practice, this means that for each sample in time, the total PINN loss is defined as follows: (1) in, The loss weights are defined as follows: (2) The goal of this method is to minimize time only when all losses corresponding to previous times have been minimized. This method minimizes the loss. It significantly improves the results of forward problems by allowing information to flow forward from the known initial conditions in a timely manner.

[0306] Since the causal training mentioned above considers all past losses, this can be computationally cumbersome when a long history exists. Furthermore, depending on the dynamical system, it may not even be necessary to look at the entire history, as the most recent past captures larger errors that have already propagated through the dynamics.

[0307] Therefore, in another implementation, the weights can be a function of the time gradient. For illustration, time... weight The time gradient of the loss function depends on at least one previous time step and optionally several (n) last time steps, for example, . (3) The function F can be configured such that it is a monotonically decreasing function of the loss at one or more past time steps and a monotonically increasing function of the time derivative of the loss at one or more past time steps. In other words, the weights can be configured to penalize previous high losses and smaller time derivatives of the loss.

[0309] In one implementation, the weights can be an exponent of a function that is indirectly proportional to the time gradient from the last time step and directly proportional to the current loss, as shown below: (4) The training loss is given by equation (1).

[0310] Historical awareness training improves the accuracy of PIML model output.

[0311] Figure 22 The accuracy (mean squared error, MSE) of the PIML model is shown as a function of time during the period of capturing the measurements. Curve 191 represents the accuracy without using history-aware training. Curve 192 represents the accuracy when using history-aware training according to equation (2), which takes into account all previous time steps. Curve 193 represents the accuracy when using history-aware training according to equation (4), where n=1. Therefore, history-aware training, which in particular relies on the time derivative of the loss function at several (but not all) previous time steps, provides an improvement over the other two techniques 191 and 192. History-aware training 192 and 193 provide an improvement over non-history-aware training 191.

[0312] Figure 23 The accuracy (mean squared error, MSE) of the PIML model trained according to equation (4) for various values ​​of n at the considered previous time points is shown. The curves show the variation over time within the time period of capturing the measurement results. Curve 191 corresponds to n=1, curve 192 corresponds to n=2, curve 193 corresponds to n=10, and curve 194 corresponds to n=100. Figure 12 As shown, training that considers a long history (n=100) performs worse than training that considers a short history (n=10).

[0313] Figure 24 This is a flowchart of a method 200 for providing a PIML model for use by a processing system 50. Method 200 can be executed automatically by a computing system 170 or by a processing system 50.

[0314] At procedure block 201, perform PIML model training. References can be used. Figure 18 , Figure 19 , Figure 20 , Figure 21 , Figure 22 , Figure 23 and Figure 24 Use any of the techniques described to perform PIML model training.

[0315] At process block 202, PIML model verification is performed. PIML model verification may include using a first parameter determined by the PIML model in a spatially resolved manner as input to a simulation technique that provides a second parameter (also determined by the PIML model) in a spatially resolved manner. PIML model verification may include comparing the spatially resolved PIML model output of the second parameter with the spatially resolved second parameter determined using the simulation technique. This facilitates verification when measurement results are unavailable or unreliable as real data for PIML model verification.

[0316] Simulation techniques that can be used to simulate physical processes in assets are commercially available. The COMSOL simulation system is an example of this. Therefore, the dynamics of asset parameters (such as thermodynamics controlled by thermal diffusion) can be modeled for PIML model validation, as explained.

[0317] Figure 25 This is a flowchart of method 210 for providing a PIML model for use by processing system 50. Method 210 can be executed automatically by computing system 170 or by processing system 50. Method 210 can be executed to implement PIML model verification 202.

[0318] At process block 211, the PIML model is used to predict the heat distribution. The predicted heat distribution (whose true data is often not fully available from measurements) can be used to simulate the system to reconstruct the temperature distribution.

[0319] At process block 212, a simulation is performed using the heat distribution output from the PIML model as input to obtain a reconstructed temperature distribution. Simulation tools are available in the art and can be used to simulate the thermodynamics within an asset. The result of the simulation is a reconstructed temperature distribution.

[0320] At process block 213, the reconstructed temperature distribution is compared with the measured temperature distribution.

[0321] This technique for performing PIML model verification addresses the challenge that, when considering constraints imposed by assets such as power transformers, it may be difficult to obtain accurate data values ​​for at least one parameter (such as heat) using conventional measuring instruments. Figure 25 The method provides an indirect way to validate the PIML model by using a portion of the PIML model output (the estimated heat function) to reconstruct the temperature distribution using any established physical simulation software (such as COMSOL) and comparing the reconstructed temperature with the measured temperature.

[0322] Figure 26The reconstructed temperature (varying over time) was compared with the temperature measured at selected points on the surface of the transformer wall. Figure 26 In the diagram, curves 221, 223, 225, 227, 229, and 231 show the measurement results. Curves 222, 224, 226, 228, 230, and 232 show the reconstructed temperatures. The location is specified in an inset plot for each box.

[0323] Figure 27 The arrangement of loc-1, loc-2, loc-3, loc-4, loc-5, and loc-6 at different positions is shown. For these positions, Figure 26 A comparison of a portion of the box wall 240 is provided.

[0324] Figure 28 The variation of the validation mean square error (MSE) 250 over the training period is shown.

[0325] Therefore, as the data provided demonstrates (see, for example) Figure 26 and Figure 28 This achieved good prediction accuracy.

[0326] For the data disclosed in this paper, the training data was captured using a thermal camera with a chip exceeding 1000 × 1000 pixels, specifically a series of 200 images captured over a period of several minutes during load current variations. The training data was captured on a power transformer with a tank filled with insulating oil.

[0327] These processing systems and methods are operable for processing measurement results captured using a thermal camera. The resolution of the thermal camera can be, for example, at least 500 × 500 pixels, at least 800 × 800 pixels, or at least 1000 × 1000 pixels. Thermal images can be captured over a period of several minutes, wherein at least 50, at least 100, or at least 150 images are captured and processed.

[0328] To illustrate, Figure 22 , Figure 23 , Figure 26 and Figure 28 The data shown was obtained using a thermal camera with over 1000 × 1000 pixels, where over 150 images were captured and processed within a few minutes.

[0329] However, it has been surprisingly established and demonstrated that these techniques (i.e., processing systems and methods) remain operational even when using much sparser data. To demonstrate this, the performance of the PIML model was examined for the case of predicting spatiotemporal variations in temperature and / or heat using only a subset of pixels from a thermal camera.

[0330] More specifically, the trained PIML model was tested using inputs including only 1 / 10 and 1 / 2000 pixels of measurements from a thermal camera. As the tests show, performance remained good even under sparse data conditions (where only a subset of pixels (1 / 10 and 1 / 2000 pixels) of the thermal camera were used to generate the input for PIML model 80). Therefore, the technique disclosed in this paper provides robustness under sparse data conditions.

[0331] PIML models can be used to generate outputs that can take various forms.

[0332] Figure 29 This is a flowchart of method 260. Method 260 can be executed automatically by the processing system 50 alone or in association with a control system 49 communicatively coupled to the processing system 50.

[0333] At process block 261, processing system 50 determines which action to perform based on at least one defined parameter. The action to perform may be determined based on spatial and / or temporal variations of the defined at least one parameter. Determining the action to perform may include determining whether unacceptable localized damage exists in the asset based on spatiotemporally varying temperature and / or heat distribution.

[0334] At process block 262, the action defined at process block 261 is performed. Performing the action may include outputting any one or any combination of a warning, alert, or alarm. Alternatively or additionally, performing the action may include the operation of major systems affecting the power system, including asset 20, performed by processing system 50 or by control system 49 communicatively coupled to processing system 50.

[0335] Figure 30 A schematic representation of system 10 according to a further embodiment is shown. System 10 includes measuring instruments comprising a plurality of sensors 271 (such as thermocouples) arranged and operable to sense temperatures at multiple locations on one or more sidewalls 31, 32 and / or cover 35 of housing 30. Processing system 50 is operable to process data 58 including the measurement results captured by the plurality of sensors 271.

[0336] As already explained, the processing system 50 can operate to trigger an action based on at least one determined parameter. The processing system 50 can be integrated into or communicatively coupled to the automation control system of the power system. (Reference) Figure 31 , Figure 32 , Figure 33 , Figure 34 and Figure 35 Let's explain the exemplary configuration.

[0337] Figure 31 , Figure 32 , Figure 33 and Figure 34 A schematic representation of system 10, including an electrical system, is shown. The electrical system includes a main system, which may include asset 20. The main system may further include components suitable for protecting asset 20, such as circuit breakers or switchgear 281. The electrical system may further include an automation control system 290. The automation control system 290 may include multiple intelligent electronic devices (IEDs). The automation control system 290 may include a communication system 300 that allows the intelligent electronic devices to communicate with each other and / or with control system 49. HMI 299 may be coupled to or included in control system 49. Communication system 300 may include communication infrastructure, such as a communication switch 301, for enabling communication between IEDs. As previously explained, system 10 includes a measuring instrument 41 capable of operating to capture measurement data of asset 20, such as thermal images or other thermal measurement results. System 10 further includes a processing system 50, which is operable to process the measurement results to determine one or more parameters in a spatially resolved manner, in this sense determining one or more parameters for multiple locations on the box of asset 20.

[0338] Figure 31 System 10 is shown, comprising a first IED 291 operable to perform control and / or monitoring functions on asset 20. Asset 20 may be a power transformer or a shunt reactor, but is not limited thereto. System 10 includes a second IED 292 operable to trip switchgear 281. System 10 may be operable to cause processing system 50 to generate an output that causes at least one of IEDs 291, 292 to perform an operation in response to the output of processing system 50, wherein the output of processing system 50 depends on at least one determined parameter. Alternatively or additionally, system 10 may be operable to cause processing system 50 to generate an output that causes warnings, alerts, alarms, or other status information related to asset 20 to be provided via HMI 299, wherein the warnings, alerts, alarms, or other status information depends on at least one determined parameter.

[0339] Figure 32 It shows the Figure 30 System modifications. In Figure 31 In this system, at least one IED 291 of the system 10 includes a processing system capable of operating as disclosed herein.

[0340] The techniques disclosed herein are particularly applicable to power systems that include distributed energy resources or other renewable energy sources and / or time-coupled devices such as vehicle charging equipment.

[0341] Figure 33 A system 10 is shown that includes a processing system 50, wherein the power system includes a main system device 280, which includes assets 20 and at least one of renewable energy 282 and electric vehicle charging equipment 283.

[0342] The techniques disclosed herein are also particularly applicable to power systems, including high-voltage direct current (HVDC) systems.

[0343] Figure 34 System 10, including processing system 50, is shown. Major system equipment 280 includes a power transformer 20 coupled to AC system 284 and inverter / converter 286. Major system equipment 280 includes HVDC system 285. HVDC system 285 includes DC transmission lines or cables implemented redundantly via DC transmission lines or cables 288, 288'. HVDC system 285 includes protective devices for lines or cables 288, 288', such as disconnect switches 287, 287'.

[0344] System 10 includes measuring instrument 41. Processing system 50 is operable to generate an output based on the output of processing system 50 that causes at least one IED of automation control system 290 to operate. The output of processing system 50 depends on predicted parameter values. System 10 is operable to operate at least one of IED 291 associated with power transformer 20, IED 296 associated with converter / inverter 286, and IED 297, 298 associated with disconnect switches 287, 287' in a manner that affects major system equipment 280, based on the output of processing system 50.

[0345] Figure 35 A schematic representation of system 10, including processing system 50, is shown. System 10 further includes SCADA system 310, which includes control subsystem 311. Control subsystem 311 is operable to control main system equipment 312 based on output generated by processing system 50. Control subsystem 311 can operate to change the operation of asset 20 or other main system equipment related to the operation of asset 20, for which processing system 50 determines at least one parameter. For illustration, the identification of localized losses in asset 20 by processing system 50 may cause control subsystem 311 to change the load current and / or load power of the asset, change the tap changer position, or perform another control action suitable for mitigating a potentially critical situation.

[0346] The processing systems, methods, and systems according to the embodiments achieve various effects and advantages. These processing systems and methods provide enhanced techniques for determining at least one parameter for an asset in a spatially resolved manner using data-driven processing techniques in which physical knowledge can be embedded. These techniques are capable of manipulating input data for designing physical information data-driven processing techniques in such a way that measurements that do not reflect the physical relationships embedded in the physical information data-driven processing techniques are discarded or replaced.

[0347] These processing systems and methods allow the use of PIML models to determine temperature and / or heat, thus enabling the determination of temperature and / or heat in areas where measurements on the side walls and / or lid of the enclosure are unreliable or unavailable.

[0348] Although embodiments have been described in detail with reference to the accompanying drawings, various modifications may be implemented in other embodiments. This is for illustrative purposes only and not for limitation. • While embodiments have been described in connection with assets including power transformers or shunt reactors with enclosures, these techniques can also be used in connection with other equipment, such as measuring transformers.

[0349] • While an embodiment in which the processing system 50 is included in or connected to a SCADA system via a communication interface has been described, the processing system 50 may also be used in other use cases, such as for performing offline asset analysis.

[0350] • While embodiments in which a heat diffusion equation is embedded in a PIML model and trained to predict temperature and / or heat, as well as a measure of temperature and / or heat uncertainty, have been described, the techniques disclosed herein are not limited to this use case. For illustration, the techniques disclosed herein can be used in connection with PDEs other than heat diffusion, particularly with other physical parameters of an asset that must satisfy a PDE that correlates the first-order partial time derivative and the second-order spatial derivative (such as a 2D or 3D Laplace operator) with each other and / or with another parameter.

[0351] • The techniques disclosed herein can be used to predict both the mean and uncertainty measure (such as variance or logarithm of variance) of the probability distribution of at least one parameter in a spatially resolved manner. Therefore, the techniques disclosed herein can be used with the processing techniques disclosed in application EP23209949.9, filed November 15, 2023, entitled “METHODAND PROCESSING SYSTEM FOR USE WITH AN ASSET OF AN ELECTRIC POWER SYSTEM”.

[0352] The embodiments may be used in connection with grids with renewable energy penetration (such as grids that include renewable energy systems (such as DERs)) and / or with grids that have HVDC systems, but are not limited thereto.

[0353] This specification and accompanying drawings, which illustrate various aspects and embodiments of the invention, should not be considered as limiting the scope of the claims. In other words, while the invention has been illustrated and described in detail in the accompanying drawings and the foregoing specification, such illustrations and descriptions should be considered illustrative rather than restrictive. Various mechanical, compositional, structural, electrical, and operational changes can be made without departing from the spirit and scope of this specification and the claims. In some instances, well-known circuits, structures, and techniques have not been shown in detail to avoid obscuring the invention. Therefore, it will be understood that changes and modifications can be made by those skilled in the art within the scope and spirit of the following claims. In particular, the invention covers other embodiments having any combination of features from the different embodiments described above and below.

[0354] This disclosure also covers all other features shown individually in the drawings, although these features may not be described in the preceding or following description. Furthermore, single alternatives to the embodiments described in the drawings and specification, and single alternatives to their features, may be excluded from the subject matter of the invention or the disclosed subject matter. This disclosure includes subject matter constituted by the features defined in the claims or embodiments, as well as subject matter including said features.

[0355] The term "comprising" does not exclude other elements or process blocks, and the indefinite articles "an" or "a" do not exclude a plural. A single unit or process block can perform the function of multiple features recited in the claims. The simple fact that certain measures are recited in mutually different dependent claims does not imply that a combination of these measures cannot be used advantageously. Components described as coupled or connected can be directly electrically or mechanically coupled, or they can be indirectly coupled via one or more intermediate components. No reference numerals in the claims should be construed as limiting the scope.

[0356] Machine-readable instruction code can be stored / distributed on a suitable medium (such as an optical storage medium or solid-state medium provided with or as part of other hardware), but can also be distributed in other forms (such as via a wide area network or other wired or wireless communication system). Furthermore, machine-readable instruction code can also be a data structure product or signal used to embody a specific method (such as the method according to an embodiment).

Claims

1. A method for use in connection with an asset (20) of a power system (10), the method comprising: Data is received at at least one interface (51) of the processing system (50), the data including measurement results (91) at or near several different measurement locations at or near the asset (20). The determination of at least one parameter of the asset (20) is performed by at least one processing circuit (60) of the processing system (50), including: At least based on the measurement results (91), input data (65) for physical information data-driven processing techniques (63; 120; 130) is generated, including identifying a subset of reliable measurement locations of the measurement results (91) and using the measurement results (91) obtained at the identified subset to generate the input data (65), and The physical information data-driven processing technique (63; 120; 130) is applied to the input data (65) to determine the at least one parameter, and The processing system (50) generates the output (59) based on the at least one parameter.

2. The method according to claim 1, wherein, The measurement result (91) includes a spatially resolved temperature measurement result (91), wherein identifying the subset of measurement locations includes at least one of the following operations: performing image segmentation based on a temperature gradient determined according to the temperature measurement result (91) and / or based on the construction data of the asset (20), wherein the at least one parameter indicates electrical power loss.

3. The method according to claim 2, wherein, The spatially resolved temperature measurement result (91) includes the pixel value or voxel value of at least one thermal image of at least one wall (31, 32, 35) of the box (30) of the asset (20).

4. The method according to any one of the preceding claims, wherein, The physical information data-driven processing technology (63; 120; 130) includes at least one physical information machine learning PIML model (120; 130).

5. The method according to claim 4, wherein, The at least one PIML model (120; 130) has at least one output, which is operable to provide the temperature and heat determined by the at least one PIML model (120; 130).

6. The method according to claim 5, wherein, The at least one PIML model (130) includes: a first plurality of neural network NN layers (131), the first plurality of NN layers having a first output (133) and a first input (131) operable to receive the input data (65); a second plurality of NN layers (134), the second plurality of NN layers having a second output (136) operable to provide the temperature and a second input (135) operable to receive the first output data provided by the first output (133); and a third plurality of NN layers having a third plurality of NN layers (137), the third plurality of NN layers having a third output (139) operable to provide the heat and a third input (138) operable to receive the first output data provided by the first output (133).

7. The method according to any one of claims 4 to 6, wherein, The at least one PIML model (120; 130) incorporates a partial differential equation for thermal diffusion (PDE).

8. The method according to any one of claims 4 to 8, wherein, The PIML model (120; 130) includes at least one output layer, which is operable to enforce the positivity of the temperature and heat determined by the at least one PIML model (120; 130).

9. The method according to any one of the preceding claims, wherein, The method further includes training the at least one physical information data-driven processing technique (63; 120; 130) by the processing system (50) or by a computing system (170) separate from the processing system (50), wherein the training includes historical perception training. Optionally, the history-aware training is performed using a loss function to be maximized or minimized, wherein the loss function has at least one of the following: This includes a weighting factor that depends on the training loss at past moments of the training. Introduce time gradient regularization; The training loss is penalized based on the derivative of the loss function.

10. The method of claim 9, further comprising interfaced by the processing system (50) or the computing system (170) with an asset simulator (169) simulating the thermal behavior of the asset (20) to perform model verification on the at least one physical information data-driven processing technique (63; 120; 130).

11. The method according to any one of the preceding claims, wherein, The assets (20) include transformers and / or shunt reactors.

12. A method for controlling an electric power system (10), the method comprising: The processing system (50) performs the method according to any one of the preceding claims to determine the losses in the assets (20) of the power system (10), and The processing system (50) or the control system (49) operatively coupled to the processing system (50) performs at least one control operation based on the output (59).

13. A method for providing a physical information machine learning PIML model (120; 130), the method comprising: Training data (178) is obtained by the computing system (170), which is based on measurements (91) captured on the box walls (31, 32, 35) of the asset (20). The computing system (170) trains the PIML model (80) using the training data (178), wherein the training includes history-aware training, in which the loss function at each training time depends on the loss function at at least one past training time of the training, and The PIML model data obtained during the training is provided for use in the method as described in any of the preceding claims.

14. The method according to claim 13, wherein, The historical perception training depends on the time derivative of the loss function, and the method further includes PIML model validation, wherein the PIML model validation includes using spatially resolved thermal predictions provided by the PIML model to determine the spatially resolved simulated temperature distribution by the computing system (170) or a simulator (169) connected to the computing system (170) via a communication interface, and comparing the spatially resolved simulated temperature distribution with spatially resolved temperature measurement results.

15. A processing system (50) for use in association with an asset (20) of a power system (10), the processing system (50) comprising: At least one interface (51) is operable to receive data (58), the data including measurement results (91) at or near the asset (20). At least one processing circuit (60) is operable to perform the determination of at least one parameter of the asset (20), wherein the at least one processing circuit (60) is operable to perform the determination of at least one parameter of the asset (20). At least based on the measurement results (91), input data (65) for physical information data-driven processing techniques (63; 120; 130) is generated, including identifying a subset of reliable measurement locations of the measurement results (91) and using the measurement results (91) obtained at the identified subset to generate the input data (65), and The physical information data-driven processing technique (63; 120; 130) is applied to the input data (65) to determine the at least one parameter. The processing system (50) is operable to generate an output (59) based on the at least one parameter.

Images