System and method for performing meter-to-transformer mapping in an electrical distribution network

Abstract

There is provided a computer-implemented method for performing meter-to-transformer mapping in an electrical distribution network, comprising: obtaining (M1) measurements from an electricity meter at a plurality of time points; generating (M2) a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; providing (M3) the sequence of meter data to a machine learning model 100; generating (M4), by the machine learning model, ranking scores for each of a plurality of transformers; and selecting one of the plurality of transformers which has the highest ranking score for mapping with the electricity meter.

Description

[0001]SYSTEM AND METHOD FOR PERFORMING METER-TO-TRANSFORMER MAPPING IN AN ELECTRICAL DISTRIBUTION NETWORK Technical Field The present disclosure relates to a system and a method for performing meter-to- transformer mapping in an electrical distribution network. More particularly, but not exclusively, the present disclosure relates to a computer-implemented method of selecting a suitable / compatible transformer for supplying power to a given smart electricity meter and its load, by using a machine learning model which processes time- series meter data from the given smart electricity meter. Background In electrical distribution networks, accurate mapping of smart electricity meters (hereinafter “meters”) to distribution transformers (hereinafter “transformers”) is beneficial for monitoring load distribution and analysing energy consumption. The distribution transformers typically perform step-down voltage conversion and also manage power distribution to various downstream sites (each associated with an electricity meter). The meter-to-transformer mapping involves associating individual electricity meters with respective transformers such that the transformers supply electricity to the meters (and their load) within the distribution grid. Traditional mapping methods often rely on static geographical information systems (GIS) and / or manual analysis, which can be time-consuming and error-prone, leading to issues such as incorrect transformer loading, reduced transformer lifespan, and increased troubleshooting time. Further, the traditional methods are increasingly inadequate with the growing complexity of modern power grids. As such, there are challenges in ensuring accurate meter-to-transformer mappings, particularly in complex network topologies with multiple feeders and service locations. Further, as the power load of a meter increases or fluctuates, the phase and voltage characteristics of the meter may change. Examples of such a meter include a poly-phase meter and a three-phase meter. Polyphaser meters are designed to handle more than one phase of electrical power, and are typically used in higher power applications such as industrial, commercial, or large residential setups. Polyphaser meters can measure power across multiple phases A three-phase meter is designed to measure electrical consumption across all three phases (Phase A, Phase B, and Phase C), but may work in a single-phase system depending upon the specific meter design. For a poly-phase or a three-phase meter, the meter may switch from operating in the single-phase mode to the double-phase or even three-phase mode, in response to the rising of its power load. The increase of power load may be caused by new appliance(s) (e.g., an EV charging point or large equipment) installed at a downstream of the meter. Consequently, the transformer originally mapped to the meter before the increase of the power load (e.g., a single-phase transformer) may be no longer suitable for supplying power to the meter and its load. Continuing to use the original transformer without updating the meter-to-transformer mapping may potentially lead to power outage or other disruptions to the electrical distribution network. Traditional methods often lack the capability to incorporate real-time meter data so as to update the meter-to-transformer mapping. Further, high similarity in consumption patterns among meters connected to different transformers makes it difficult to distinguish between them using traditional methods. For example, meters that experience significant load variations may show voltage characteristics that resemble those of meters connected to different transformers. This further complicates the meter- to-transformer mapping process. It is an object of the present disclosure, among others, to provide an improved meter-to- transformer mapping method, which solves the problems associated with traditional mapping methods, whether identified herein or otherwise. Summary According to a first aspect of the present disclosure, there is provided A computer- implemented method for performing meter-to-transformer mapping in an electrical distribution network, comprising: obtaining measurements from an electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; providing the sequence of meter data to a machine learning model; generating, by the machine learning ranking scores for each of a plurality of transformers; and selecting one of the plurality of transformers which has the highest ranking score for mapping with the electricity meter. Advantageously, the method of the first aspect improves the accuracy of meter-to- transformer mapping as compared to traditional methods, and also allows the meter-to- transformer mapping to be updated in nearly real-time based upon the latest consumption pattern of the electricity meter as reflected in the sequence of meter data. It would be understood that the machine learning model may be any temporal sequence model capable of handling time-series input data, including but not limited to recurrent neural networks (RNNs), Gated Recurrent Units (GRUs) or Transformer-based deep- learning architecture. It would further be understood that the machine learning model is a pre-trained machine learning model. The electricity meter comprises a smart electricity meter. The computer-implemented method may further comprise: reconfiguring the electrical distribution network such that the selected one of the plurality of transformers supplies power to the electricity meter. The meter data within each entry of the sequence may comprise: first data characterising a voltage measurement of the electricity meter, second data characterising a current measurement of the electricity meter and third data characterising a power measurement of the electricity meter. The electricity meter may have more than one active phase. At least one of the first to third data may comprise an aggregate value of data characterising a corresponding measurement of the electricity meter in each of the more than one active phase. The machine learning model may comprise a recurrent neural network. The recurrent neural network may comprise a Long short-term memory, LSTM, model. The machine learning model may have an encoder-decoder neural network architecture, which comprises an encoder, a self- model and a decoder. The encoder may comprise a recurrent neural network which is configured to generate a sequence of hidden states based upon the sequence of meter data. The self-attention model may be configured to generate a context vector based upon the sequence of hidden states. The decoder may be configured to generate the ranking scores for each of the plurality of transformers based upon the context vector. The decoder may be configured to apply weights to the context vector and also may comprise a function layer which is configured to generate the ranking scores for the plurality of transformers based upon the weighted context vector. The function layer may be configured to normalize the weighted context vector to a probability distribution which represents the ranking scores of the plurality of transformers. The function layer may comprise a softmax layer. A length of the sequence of meter data may be longer than or equal to a number of the plurality of transformers. The electricity meter may comprise a multi-phase electricity meter or a three-phase electricity meter. Alternatively, the electricity meter may be any other configuration of electricity meter capable of measuring electrical parameters across multiple phases. A number of active phase(s) of the electricity meter may be changeable in response to a variation in power load of the electricity meter. It would be understood that, after the number of active phase(s) of the electricity meter changes, a different one of the transformers may achieve the highest ranking score and thus be selected for mapping with the electricity meter. The electricity meter may be a first electricity meter, and the method may further comprise: generating a training dataset, wherein the training dataset comprises, for each of a first plurality of electricity meters: a sequence of meter data generated based upon measurements obtained from the respective electricity meter at a plurality of time points, and target label data indicative of a transformer which is one of the plurality of transformers; and training the machine learning model based upon the training dataset. The first electricity meter may be one of first plurality of electricity meters. Generating the training dataset may comprise: obtaining measurements from a second plurality of electricity meters at a plurality of time points, wherein the first plurality of electricity meters are included within the second plurality of electricity meters; obtaining information indicative of transformers which are associated with respective ones of the second plurality of electricity meters; and one or more of the following steps: calculating a distance between each of the second plurality of electricity meters and its associated transformer, and if the calculated distance for a first one of the second plurality of electricity meters is longer than a distance threshold, excluding data associated with the first one of the second plurality of electricity meters from the training dataset; and determining whether any of the obtained measurements comprise anomalous measurements, and if the anomalous measurements were obtained from a second one of the second plurality of electricity meters, excluding data associated with the second one of the second plurality of electricity meters from the training dataset. It will be understood that the remaining ones of the second plurality of electricity meters form the first plurality of electricity meters. Determining whether any of the obtained measurements comprise anomalous measurements may be based upon variations between the measurements obtained from a same electricity meter. Generating the training dataset may comprise generating the target label data based upon the information indicative of transformers which are associated with respective ones of the second plurality of electricity meters. It would be understood that the target label data is indicative of a transformer which is determined as being suitable for supplying power to the respective electricity meter. The transformer indicated by the target label data may be identical to or may be different from a transformer which was historically connected to the respective electricity meter at a time when the measurements was obtained from the respective electricity meter. Generating the training dataset may comprise: obtaining geographical locations of the first plurality of electricity meters; and identifying a list of available transformers which are located in a neighbourhood of the first plurality of electricity meters, wherein the plurality of transformers are included the list of available transformers. The plurality of transformers may be identical to, or may be a subset of the list of available transformers. Generating the training dataset may comprise: organising at least some of the first plurality of electricity meters which have highly correlated measurements into a sub- group, wherein the obtained information for the electricity meters within the sub-group is indicative of a first transformer; calculating a correlation coefficient of measurements between a second electricity meter of the first plurality of electricity meters and an electricity meter of the sub-group, wherein the second electricity meter is not included within the sub-group; and if the correlation coefficient is higher than a correlation threshold, generating or modifying the target label data for the second electricity meter so as to indicate the first transformer. The obtained information may indicate that a second transformer is associated with the second electricity meter. Generating the training dataset may further comprise: determining whether the second transformer is within the list of available transformers; wherein calculating the correlation coefficient of measurements between the second electricity meter and an electricity meter of the sub-group is performed in response to determining that the second transformer is not within the list of available transformers. The correlation threshold may be an average value of the correlation coefficients between the electricity meters of the sub-group. Further or alternatively, the correlation threshold may be dynamically adjusted based on historical data trends or user-specified input. The first plurality of electricity meters may be within a geographical region, and the method may further comprise: determining whether a location of the first electricity meter is within the geographical region, wherein providing the sequence of meter data to the machine learning model is performed if but only if the location of the first electricity meter is within the geographical region. The method may further comprise: whether any of the second plurality of electricity meters is outside the region; and if it is determined that a third one of the second plurality of electricity meters is outside the geographical region, excluding data associated with the third one of the second plurality of electricity meters from the training dataset. According to a second aspect of the present disclosure, there is provided a system for performing meter-to-transformer mapping in an electrical distribution network, comprising: a processor; and a memory coupled to the processor storing instructions, which when executed by the processor, cause the processor to perform steps comprising: obtaining measurements from an electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; providing the sequence of meter data to a machine learning model; receiving from the machine learning model ranking scores for each of a plurality of transformers; and selecting one of the plurality of transformers which has the highest ranking score within the plurality of transformers for mapping with the electricity meter. The memory may further store the machine learning model. Alternatively, the machine learning model may be stored in a remote device which is in communication with the system. According to a third aspect of the present disclosure, there is provided a method of training a machine learning model for performing meter-to-transformer mapping in an electrical distribution network, comprising: generating a training dataset, wherein the training dataset comprises, for each of a first plurality of electricity meters: a sequence of meter data generated based upon measurements obtained from the respective electricity meter at a plurality of time points, and target label data indicative of a transformer which is one of a plurality of transformers located in a neighbourhood of the first plurality of electricity meters, and wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; and training the machine learning model based upon the training dataset. In the target label data, the ranking score of the transformer indicated by the target label data may be 100% and the ranking scores of other transformers are all zero. According to a fourth aspect of the present disclosure, there is provided a method of obtaining a training dataset for training a machine learning model for performing meter- to-transformer mapping in an electrical distribution network, comprising, for each of a first plurality of electricity meters: obtaining measurements from the respective electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; and obtaining target label data indicative of a transformer which is one of a plurality of transformers located in a neighbourhood of the first plurality of electricity meters; wherein the training dataset comprises the sequence of meter data and the target label data of each of the first plurality of electricity meters. The use of “based upon” is meant to be open and inclusive, in that a process, step, calculation, or other action “based on” one or more recited conditions or values may, in practice, be based on additional conditions or values beyond those recited. It would also be understood that the terms “first”, “second” and “third” are simply used in the present disclosure to label the relevant elements for the ease of description, and do not imply any limitations to the sequence or the total number of the relevant elements unless clearly indicated by the context. Where appropriate any of the optional features described above in relation to one of the aspects of the present disclosure may be applied to another one of the aspects of the disclosure. In particular, the optional features described above in relation to the first aspect may be equally applied to any one of the second to the fourth aspects of the disclosure. Brief Description of the Drawings In order that the disclosure may be more understood, a number of embodiments of the disclosure will now be described, by way of example, with reference to the accompanying drawings, in which: Figure 1 schematically illustrates processing steps of a method of generating a training dataset for training a machine learning model for performing / updating meter-to- transformer mapping in an electrical distribution network, according to an aspect of the present disclosure; Figure 2 schematically illustrates an example architecture of a machine learning model for performing / updating meter-to-transformer mapping in an electrical distribution network, according to an aspect of the present disclosure; Figure 3 schematically illustrates a Long Short-Term Memory (LSTM) unit for use in the machine learning model of Figure 2; Figure 4 shows the calculated loss values after each iteration during a training process of the machine learning model of Figure 2; Figure 5 shows the results of meter-to-transformer mapping as generated by the trained machine learning model of Figure 2; Figure 6 schematically illustrates processing steps of a method of performing / updating meter-to-transformer mapping in an electrical distribution network, according to an aspect of the present disclosure; Figure 7 schematically illustrates a modular structure of a system for performing / updating meter-to-transformer mapping in an electrical distribution network, according to an aspect of the present disclosure; Figure 8 schematically illustrates processing steps of a method of training a machine learning model for performing / updating meter-to-transformer mapping in an electrical distribution network, according to an aspect of the present disclosure; Figure 9 schematically illustrates steps of a method of obtaining a training dataset for training a machine model for performing / updating meter-to- transformer mapping in an electrical distribution network, according to an aspect of the present disclosure. In the figures, like parts are denoted by like reference numerals. It will be appreciated that the drawings are for illustration purposes only and are not drawn to scale. Detailed Description of the Preferred Embodiments Referring to Figure 1, there is provided a computer-implemented method of generating a training dataset for training a machine learning model for the purpose of performing or updating meter-to-transformer mapping in an electrical distribution network. The method is performed by a computing apparatus / system with a network interface for receiving / sending data from / to external devices (e.g., smart electricity meters). At step S1, measurements (which may also be referred to as “readings”) are obtained from a plurality of smart electricity meters. The obtained measurements may include voltage, current and power measurements. For a meter with more than one active phase, the measurements from each active phase may be obtained. For each active phase of each meter, the measurements are obtained at a series of time points (e.g., once every 30 minutes). The time interval between adjacent time points may vary depending on the network configuration and the specific requirements of the analysis. The readings of one meter may be obtained at the same or different time points as / from the readings of another meter. At step S2, the measurements are pre-processed. In an example, all of the measurements are transformed into per-unit values, according to Equation (1) below: (1) Different base values may be used for transforming voltage, current and power measurements. For example, the base voltage value may be 120V or 240V, the base current value may be 5A or 10A, and the base power value may be 100kW. Other base values are possible. It would be that a single base voltage value shall be used for all of the voltage measurements from all meters. Similarly, a single base current value and a single base power value shall be used for current and power measurements from all of the meters. Converting meter readings into per-unit values simplifies the comparison across different meters connected to different transformers, regardless of their power ratings. It normalises voltage, current, and power measurements and allows consistent analysis, thereby making the following processing steps easier. Pre-processing the measurements may also include calculating variations (delta values) of the measurements for each meter between adjacent time points. The delta values indicate the short-term fluctuations or changes in key electrical parameters like voltage, current, or power at a given meter within a short interval. If the measurements were obtained once every 30 minutes at step S1, then the short interval may be 30 minutes, 1 hour or longer. The delta values may be calculated based upon the per-unit values. The delta values may be used at step S5 of Figure 1 which is described below. At step S3, information indicative of transformers which are associated with respective ones of the plurality of electricity meters is obtained. The information may be obtained from a database (e.g., a database maintained by a utility provider). In an example, at least a part of the information has been manually verified such that the transformer associated with a respective meter is indeed suitable for supplying power to the respective meter. In another example, the information may be generated by a predictive model at an earlier time. In a further example, the information may indicate transformers which were actually connected to respective meters at some point (e.g., when the measurements were obtained at step S1). In one or more of the examples, some of the information may be out-of-date or incorrect. Therefore, the information obtained at step S3 acts as a basis for generating target label data of the training dataset, and additional processing (e.g., steps S6 to S10 below) may be performed to remove errors from the obtained information, thereby enhancing the accuracy of the target label data. At step S4, a distance between each meter and its associated transformer (which is indicated by the information obtained S3) is calculated, and if the calculated distance is longer than a distance threshold, the corresponding meter and its related data will be excluded from consideration for the generation of the training dataset. In particular, the calculated distance being longer than the distance threshold means that the corresponding meter is physically too far from its associated transformer, which could indicate an incorrect initial mapping in the information obtained at step S3 or a faulty meter. The distance may be calculated based upon information indicative of geographical locations of the meters and the transformers. The utility provider typically maintains such information. At step S5, the processing determines whether any of the obtained measurements comprise anomalous measurements, and if so, excluding data associated with a meter which generates the anomalous measurements from the training dataset. Anomalous measurements refer to measurements which deviate significantly from the norm, and could indicate faulty or anomalous behaviour of the meter. In an example, the delta values generated at step S2 are used to determine whether any of the obtained measurements comprise anomalous measurements. The delta values, which represent short-term variations, can flag a meter-transformer pair that exhibits abnormal behaviour. For instance, a meter that shows frequent, high-level, variations may be connected to an underperforming transformer or be subject to irregular loads, signalling a potential misalignment between meter and transformer capacity. For the ease of description, the plurality of meters referred to at step S1 may be referred to as “a second plurality of (electricity) meters”, and the remaining meters after step S5 may be referred to as “a first plurality of (electricity) meters”, which is a sub-set of the “second plurality of meters” due to the exclusion of some meters at steps S4 and S5. The training dataset is generated based upon the first plurality of meters. Steps S6 to S10 are for generating and refining the target label data of the training dataset. At step S6, a list of available transformers in a neighbourhood of the meters are identified. This may include: obtaining geographical locations of the meters and obtaining geographical locations of available for example, based upon data provided by the utility provider. The may be defined based upon a location-based constraint and the geographical locations of the meters. The location-based constraint may be a search radius input by a user (e.g., maximum 2km from the meters), a predetermined search radius, and / or an input derived from a geospatial system or a Geographic Information System (GIS). At step S7, at least some of the meters which have highly correlated measurements are organised into a sub-group, and the information obtained at step S3 for all the meters of the sub-group is indicative of the same transformer. For the ease of description, the transformer associated with the meters of the sub-group is referred to as a “first transformer”. The expression “highly correlated” means that correlation coefficients in measurements between meters of the sub-group would be higher than a predetermined threshold. The measurements for organising some of the meters into the sub-group may comprise voltage measurements. The measurements may further comprise current measurements, power measurements and / or phase imbalance. The first plurality of meters include a particular meter which is not within the sub-group (referred to as a “second electricity meter” or a “second meter” for ease of description). The second meter is associated with a second transformer, which is different from the first transformer, according to the information obtained at step S3. The processing proceeds to step S8 where it is determined whether the second transformer is within the list of available transformers identified at step S6. If step S8 determines that the second transformer is not within the list of available transformers, the processing proceeds to step S9, where a correlation coefficient of measurements between the second meter and a meter of the sub-group is calculated. The correlation efficient may be according to any suitable correlation measures such as Pearson correlation efficient, Spearman's rank correlation coefficient, etc.. The correlation coefficient may be calculated by applying a cross correlation or a lag correlation technique. The measurements may comprise voltage measurements, and may comprise current measurements, power measurements and / or phase imbalance. At step S10, if the correlation coefficient at step S9 is higher than a correlation threshold, target label data indicative of the first transformer is generated for the second meter. In other words, the first transformer is deemed as being suitable for supplying power to the second electricity meter. The correlation threshold used at step S10 may be the same as, or different from, the correlation threshold used to determine whether the measurements of the meters within the sub-group are highly correlated. Alternatively, the target label data of all of the first plurality of meters may be initially set / generated to indicate respective transformers, which are identical to those indicated by the information obtained at step S3. This may take place at any time before S10. In that case, the target label data of the second meter would be modified at step S10 from indicating the second transformer to indicating the first transformer. Steps S7 to S10 may be repeated multiple times until the target label data of all of the meters are indicative of transformers which are included within the list of available transformers, and / or until the target label data of any sub-set of meters with highly correlated measurements indicates the same transformer. At that time, it can be said that the target label data has been finalised or refined. At step S11, a sequence of meter data is generated for each of the first plurality of meters, based upon the measurements obtained at step S1. Each of the first plurality of meters has its own sequence of meter data, and its own target label data. It would be understood that the sequence of meter data provides the input features of the training dataset, the target label data provides the output features of the training dataset. The sequence of meter data comprises a series of meter data entries indexed in time order, with each meter data entry being generated based upon the measurements obtained at a respective one of the series of time points (at which the measurements were obtained at step S1). An example of the sequence of meter data for a particular meter is shown in Table 1 below. The first column indicates the identity of the meter (e.g., “Meter_1”). The second column indicates the time points at which the meter readings were obtained at step S1. The remaining columns indicate the values of plural input features at each time point. The total number of rows in Table 1 “t”) indicates the length of the sequence of meter data. In the training dataset, the sequences of meter data for the first plurality of meters may have a fixed length (e.g., 10). The fixed-length sequences may be obtained by using a window function. Table 1 Meter_1 2023-01-010:00:00 IF1_1 IF2_1 IF3_1 .. IFk_1 Meter 1 2023-01-010:30:00 IF1 2 IF2 2 IF3 2 .. IFk 2 3 t Tabl equence of meter data. The input features include first input feature(s) characterising voltage measurements of a respective meter, second input feature(s) characterising current measurements of the meter and third input feature(s) characterising power measurements of the meter. In an example, the first input feature(s) include a voltage amplitude (V) for each phase of the meter; the second input feature(s) include an active current (IR) and a reactive current (IX) for each phase of the meter; and the third input feature(s) include an active power (P), a reactive power (Q) and a power factor (PF) for each phase of the meter. The active current (IR) is the component of the current that contributes to the real power. The reactive current (IX) is the component of the current that contributes to the reactive power. The active power (P) is the real power consumed by a load associated with the meter. The reactive power (Q) is the power that oscillates between the source and the load. The power factor (PF) is the ratio of active power to apparent power, indicating the efficiency of power usage. In the event that the meter has more than one active phase (e.g., phases A, B, C), aggregate values characterising all of the active phases are also calculated and included within the input features. For example, the first input features may also include a voltage root-mean-square (RMS) value of the voltage magnitudes across the three phases. It would be understood that other types of aggregate values (not RMS values) may be used. In an example where the meter has three active phases and the aggregate values across the three active phases are included for each of the input features described above, there may be 24 input features in (k=24). It would be understood that columns 1 and 2 of Table 1 are shown for clarity, and that data from the remaining columns of Table 1 are fed into the machine learning model for training the model. It would be appreciated that the sequence of meter data may be generated based upon other data source(s) in addition to the measurements obtained at step S1. For example, the input features may comprise further features, such as data indicative of a geographical location of the respective meter, data indicative a temperature at or in a vicinity of the electricity meter, and / or data indicative of a grid demand. The target label data may take the form of a one-dimensional data array. An example is shown in Table 2 below: Table 2 0 0 1 … 0 0 e cou s co espo o a p e ee e a ay o a so e s, c cu e all of the transformers associated with the first plurality of meters after the processing of step S10. Each data entry is either ‘1’ or ‘0’. The transformer associated with a given meter after the processing of step S10 is assigned ‘1’, with all other transformers being assigned ‘0’. The first plurality of meters may be located within a relatively small geographical region (e.g., a town, a city or a district of a city). Accordingly, the total number of transformers associated with first plurality of meters may be at a relatively low level. The training dataset is generated after step S11. It would be understood that while it is not shown in Figure 1, the values of the input features as contained in the sequence of meter data may be subject to further processing (e.g., normalisation). In addition, the dataset generated by Figure 1 may be divided into a training dataset (80%) and a validation dataset (20%). Figure 2 schematically illustrates an example architecture of a machine learning model 100 which may be used for performing the meter-to-transformer mapping. In general, the machine learning model may be any sequence model capable of handling time- series input data (i.e., the sequence of data), including but not limited to recurrent neural networks (RNNs), Gated Recurrent Units (GRUs) or Transformer-based deep- learning architecture. The architecture shown by Figure 2 is merely an example, and other architectures may be used. Referring to Figure 2, the machine learning model 100 has an encoder-decoder neural network architecture, which comprises an encoder 10, a self-attention model 20 and a decoder 30. The encoder-decoder architecture is commonly used in sequence-to- sequence tasks, such as time-series prediction. The encoder 10 comprises a RNN which is configured to generate a sequence of hidden states (h0, h1,…ht) based upon the sequence of meter data (X0, X1,…Xt). X0to Xtmay be based upon the input features in the first row to the last row of Table 1, respectively. The RNN has a loop which allows information to be passed from one step to the next. Figure 2 shows the looped diagram as well as the unrolled diagram of the RNN. In an example, the RNN is based upon a Long short-term memory (LSTM) model as shown in Figure 3, while other types of RNNs may be used. In the LSTM model, the output at each time step includes two components: • Hidden state Hi: It represents the current output of the LSTM model at time step i. It is passed on to the next time step. The hidden state Hi shown in Figure 3 corresponds to the hidden sate hi shown in Figure 2. • Cell state Ci: It is the memory of the network, which carries information forward across long sequences. It allows the LSTM to remember patterns across many time steps and prevents the vanishing gradient problem. Referring back to Figure 2, for each time step i (i=0,1…t), the LSTM model receives an input Xi and produces an output vector hi which is the hidden state. The output hi encodes the time-dependent features (e.g., voltage, current, power measurements over time) of the meter readings. These hidden states hi are used as input to the self-attention model 20 and the decoder 30. The RNN 10 is responsible for learning temporal dependencies in the input data Xi, and for exploring how the temporal dependencies (i.e., how meter readings change at previous time steps) impact the of a transformer for supplying power to a given meter. The LSTM model are well-suited for achieving this goal. In particular, the LSTM model is specifically designed to handle time-series data where the order of data points and historical dependencies matter. In the case of meter-to-transformer mapping, the electrical parameters (voltage, current and power measurements etc.) vary over time, and the LSTM model is able to capture those temporal dynamics. This helps in making accurate predictions of suitable transformers by learning patterns of meter readings from previous time steps. Further, the LSTM model is good at handling long-term dependencies. The LSTM model excels at remembering information over long sequences due to its memory cell (Ci) and gating mechanisms (forget, input, and output gates as shown in Figure 3). This allows the LSTM model to focus on relevant past events, which is crucial for understanding long-term trends in electrical data (e.g., meter readings) that might affect transformer suitability. Further still, the LSTM model is suitable for dealing with non-stationary Data. Electrical data is often non-stationary, meaning that the statistical properties of the data (e.g., mean, variance) change over time. LSTMs are robust to these variations and can adjust to the evolving patterns in the data. The meter-to-transformer mapping can be viewed as a sequence-to-sequence task (i.e., a sequence of electrical readings to a sequence of ranking scores for transformers). The LSTM model, used in the encoder-decoder framework of Figure 2, is highly effective for the sequence-to-sequence task. The self-attention model 20 (which may also be referred to as self-attention mechanism) is configured to generate a context vector 25 based upon the sequence of hidden states h0, h1,…ht. The self-attention model 20 helps the model 100 decide which time steps in the input sequence (X0, X1,…Xt) are the most important for performing meter to transformer mapping. For example, the self-attention model 20 may weigh certain time steps more heavily if they contain important voltage fluctuations that indicate a transformer change. The self-attention model 20 may further transform the high- dimensional input features (e.g., P, Q, V, IX, PF) into a lower-dimensional space that captures the essential characteristics of meter data. The self-attention mechanism 20 may involve three key steps: • Query (Q): Represents the current time step. • Key (K): Represents all time steps in the sequence. • Value (V): Represents the values of the hidden states at each time step. Self-attention calculates a weight for each hidden state hi based on its similarity to the query. The calculation may be based upon the attention formula (2) below: (2) These weights are used to create a context vector 25, which is then passed to the decoder 30. The weights represent the importance of each time step for the meter to transformer mapping. This context vector 25 is a summary of the entire input sequence, and allows the model 100 to focus on the relevant portions of the time-series input, thereby making transformer recommendations more accurate. The decoder 30 is configured to generate ranking scores for the predetermined array of transformers (referred to in relation to Table 2 above) based upon the context vector 25. The ranking scores are included within the output 35 of the decoder 35. The ranking scores may also be referred to as “suitability Score” which indicates the suitability of each potential transformer for supplying power to a given meter. The generation of the ranking scores may be done using a fully connected (dense) layer. In particular, the ranking scores may be generated according to Equation (3) below (3) After applying softmax, the suitability score of each potential transformer is in the range of 0 to 1, and all of the suitability scores add up to 1, so that they can be interpreted as probabilities. Higher scores correspond to larger probabilities. It is noted that the example target label data shown in Table 2 follows this arrangement of the suitability scores. It would be appreciated that the softmax function is just an example and can be replaced by other probability distribution functions. The entire model 100, including the encoder 10, the self-attention model 20 and the decoder 30, may be trained end-to-end. This holistic training allows the various components of the model to optimize their parameters in a coordinated manner. In particular, as described above, the components of the model 100 are interdependent. By training all of the components together, the model 100 can learn how these components should work together to optimize the final output 35. The weights used in the model 100 may be initialized randomly or using pre-trained weights (if available). During the training of the model 100, the training dataset generated by Figure 1 may be divided into mini-batches, where each batch contains multiple sequences of meter data and the corresponding target label data. Batching speeds up the training process by allowing parallel computations. The input features of the training dataset are fed into the model 100 which generates the output 35. The loss of the entire model 100 is computed based on the difference between the output 35 and the target label data of the same meter. In the example provided above where the output 35 includes probabilities scores for each candidate transformer, cross-entropy loss may be used. Once the loss is computed, backpropagation is applied to update the weights of the encoder 10, the self-attention model 20 and the decoder 30 simultaneously. The goal is to minimize the loss so that the model accurately predicts transformer suitability. The optimization algorithm (e.g., Adam or SGD) may be used to adjust the weights of all components based on the gradients during backpropagation. This process is repeated for several epochs until converges, meaning the loss stops decreasing significantly. Figure 4 shows the loss of the model 100 after each iteration of the training process. It can be seen that the loss reduces from 3.35 to 0.18 after six iterations. At that time, the training of the model may be considered as being completed. After the completion of the training process, the model 100 is evaluated using the validation dataset. The model’s ability to predict transformer suitability for unseen meters is tested to ensure it generalizes well. The trained model 100 is able to achieve an accuracy of higher than at least 95% based upon the validation dataset. Therefore, the implementation of the model 100 can perform highly accurate meter-to- transformer mapping based on real-time readings from meters, thereby dynamically recommending most suitable transformers to respective meters. The model 100 provides a relatively low-cost solution, which also improves the accuracy and reliability of meter- to-transformer mapping as compared to existing solutions. The improved meter-to- transformer mapping is beneficial for optimising load managements of the transformers, as well as reducing energy loss and improving the overall efficiency of the electrical distribution network. Figure 6 schematically illustrates processing steps of a method for performing or updating meter-to-transformer mapping in an electrical distribution network. The method is computer implemented. Performing / updating meter-to-transformer mapping means selecting a suitable / compatible transformer from a plurality of potential transformers for connection to a given electricity meter. The selected transformer may be different from a transformer which the given meter is currently connection to. At step M1, measurements are obtained from a given electricity meter (which may also be referred to as a “first electricity meter”) at a plurality of time points. Step M1 is similar to step S1 of Figure 1. The obtained measurements include voltage, current and power measurements of the given meter. If the meter has more than one active phase, the measurements from each active phase may be obtained at the plurality of time points (e.g., once every 30 minutes). The time interval between adjacent time points may suitably vary. At step M2, a sequence of meter data is based upon the measurements. The sequence of meter data is arranged in order based upon the plurality of time points, and the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points. The sequence of meter data generated at step M2 may be similar to the sequence of meter data generated at step S11, thus the characteristics of the sequence described above in relation to step S11 may also apply to step M2. More specifically, the meter data within each entry of the sequence may comprise: first data characterising a voltage measurement of the electricity meter, second data characterising a current measurement of the electricity meter and third data characterising a power measurement of the electricity meter. The second data may comprise an active current value and a reactive current value. Alternatively, the second data may comprise any other type of current measurement relevant to the operation of the electricity meter. The third data may comprise an active power value and a reactive power value. Further or alternatively, the third data further comprises a power factor value. Further or alternatively, the third data may comprise other relevant power measurements used for determining the suitability of a candidate transformer. In the event that the electricity meter has more than one active phase, at least one of the first to third data may further comprise an aggregate value (e.g., a root-mean-square value) of data characterising a corresponding measurement of the electricity meter in each active phase. At step M3, the sequence of meter data is fed to a pre-trained machine learning model, such as the model 100 trained using the training dataset generated by the method of Figure 1. At step M4, the machine learning model generates ranking scores for each of a plurality of candidate transformers. The ranking scores may be included within the output 35 of the model 100. At step M5, one of the plurality of candidate transformers which has the highest ranking score within the candidate transformers is selected for mapping with the electricity meter. It would be understood that the plurality of candidate transformers would be identical to the predetermined array of by the target label data of the training dataset as described above. Therefore, the ranking scores may be readily associated with the predetermined array of transformers, thereby indicating suitability of each candidate transformer for mapping with the electricity meter. Consequently, the identity of the transformer with the highest ranking score can be easily determined by cross- referencing with the predetermined array of transformers. Figure 5 shows the results generated by the trained machine learning model 100. The column labelled as “movable_meter_id” shows the IDs of smart electricity meters yet to be mapped. The column labelled as “proposed_transformer_id” shows the IDs of transformers selected by the model 100 at step M5 for each smart electricity meter. As described above, the first plurality of meters (from which the training dataset is generated) may be located within a relatively small geographical region (e.g., a city or a district of a city). Therefore, the trained model 100 is associated with the geographical region alone, and may ideally be used to perform map-to-transformer mapping for meters located within the same geographical region. The method of Figure 6 may include a further processing step of determining whether a location of the first electricity meter is within the geographical region. Steps M1 to M5 described above may be performed if but only if the location of the first electricity meter is within the geographical region. The geographical region may be pre-determined based on geographical or operational constraints, and / or other location-based criteria such as grid topology or energy demand distribution. The geographical region may be dynamically adjusted based upon real-time geographical or operational constraints. The method may include an optional step M6. At step M6, the electrical distribution network is automatically reconfigured such that the selected one of the plurality of transformers supplies power to the electricity meter. In other words, the selected one of the plurality of transformers may be automatically connected to the electricity meter and its load. This may be done through the use of smart grid infrastructure. Some modern electrical systems, especially in smart grids, can remotely control which transformers supply power to specific meters using advanced distribution management systems (ADMS) and SCADA (Supervisory Control and Data Acquisition). Alternatively, the connection between the recommended transformer (i.e., the transformer with the highest ranking score) and the meter may be manually by a technician or in a semi- automated manner. Figure 7 shows a system 50 for performing meter-to-transformer mapping in an electrical distribution network. The system 50 comprises a processor 60, a memory 70 and a network interface 80 for establishing data communication between the system and external devices / systems. Examples of the system 50 includes a portable computing device, a server, and / or any suitable computing devices. The memory 70 is operatively coupled to the processor 60 and the network interface 80. The network interface 80 allows the apparatus 50 to communicate with smart electricity meters and / or remote servers. The memory 70 stores instructions executable by the processor. In the event that the memory 70 also stores the trained machine learning model (e.g., such as the model 100 trained using the training dataset generated by the method of Figure 1), the instructions stored by the memory 70 may cause the processor 60 to perform the processing steps of Figure 6. Alternatively, the memory 70 does not store the trained machine learning model, and the trained machine learning model is stored in a remote device in communication with the system via the network interface 80. In that case, the instructions stored by the memory 70 may cause the processor 60 to perform the processing steps of M1 to M3, a step of receiving from the machine learning model ranking scores for each of a plurality of transformers and step M5 (and optionally step M6). The remote device storing the trained machine learning model may receive the sequence of data and directly provide the identity of the transformer with the highest ranking score to the system 50. If that is the case, the instructions stored by the memory 70 may cause the processor 60 to perform the processing steps of M1 to M3, and a step of receiving from the machine learning model an identity of one of a plurality of transformers which has the highest ranking score within the plurality of transformers. Step M6 may optionally be performed by the processor 60 as well. Figure 8 schematically illustrates steps of a method of training a machine learning model (e.g., the model 100) for meter-to-transformer mapping in an electrical distribution network. At step T1, a training dataset is generated. The training dataset comprises, for each of a plurality of electricity meters: a sequence of meter data generated based upon measurements obtained from the respective electricity meter at a plurality of time points, and target label data indicative of a transformer which is one of a plurality of transformers located in a neighbourhood of the plurality of electricity meters. The sequence of meter data is arranged in time order based upon the plurality of time points, and the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points. In an example, the training dataset may be generated by the method of Figure 1. All the features described above in relation to Figure 1 also apply to step T1. At step T2, the machine learning model is trained based upon the training dataset. The model may be trained end-to-end using backpropagation as described above. Figure 9 schematically illustrates processing steps of a method of obtaining a training dataset for training a machine learning model (e.g., the model 100) for performing meter- to-transformer mapping in an electrical distribution network. The processing steps O1 to O3 described below are performed for each of a first plurality of electricity meters. At step O1, measurements are obtained from the respective electricity meter at a plurality of time points. Step O1 is similar to step S1 of Figure 1. At step O2, a sequence of meter data is generated based upon the measurements. The sequence of meter data is arranged in time order based upon the plurality of time points, and the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points. Step O2 is similar to step S11 of Figure 1. At step O3, target label data is obtained. The target label data is indicative of a transformer which is one of a plurality of transformers located in a neighbourhood of the first plurality of electricity meters. In an the target label data is obtained by steps S3 and S6 to S10 of Figure 1. The dataset comprises the sequence of meter data and the target label data of each of the first plurality of electricity meters. In the methods described above in relation to Figure 1, 6, 8 and 9, the processing steps may be performed in an order which is different from the sequence described. The terms “having”, “containing”, “including”, “comprising” and the like are open and the terms indicate the presence of stated structures, elements or features but not preclude the presence of additional elements or features. The articles “a”, “an” and “the” are intended to include the plural as well as the singular, unless the context clearly indicates otherwise. Although the disclosure has been described in terms of preferred embodiments as set forth above, it should be understood that these embodiments are illustrative only and that the claims are not limited to those embodiments. Those skilled in the art will be able to make modifications and alternatives in view of the disclosure which are contemplated as falling within the scope of the appended claims. Each feature disclosed or illustrated in the present specification may be incorporated in the disclosure, whether alone or in any appropriate combination with any other feature disclosed or illustrated herein.

Claims

CLAIMS:

1. A computer-implemented method for performing meter-to-transformer mapping in an electrical distribution network, comprising: obtaining measurements from an electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; providing the sequence of meter data to a machine learning model; generating, by the machine learning model, ranking scores for each of a plurality of transformers; and selecting one of the plurality of transformers which has the highest ranking score for mapping with the electricity meter.

2. The computer-implemented method of claim 1, further comprising: reconfiguring the electrical distribution network such that the selected one of the plurality of transformers supplies power to the electricity meter.

3. The computer-implemented method of claim 1 or 2, wherein the meter data within each entry of the sequence comprises: first data characterising a voltage measurement of the electricity meter, second data characterising a current measurement of the electricity meter and third data characterising a power measurement of the electricity meter.

4. The computer-implemented method of claim 3, wherein the electricity meter has more than one active phase, and wherein at least one of the first to third data comprises an aggregate value of data characterising a corresponding measurement of the electricity meter in each of the more than one active phase.

5. The computer-implemented method of any preceding claim, wherein the machine learning model comprises a recurrent neural network.

6. The computer-implemented method of claim 5, wherein the recurrent neural network comprises a Long short-term memory, LSTM, model.

7. The computer-implemented of any preceding claim, wherein the machine learning model has an encoder-decoder neural network architecture, which comprises an encoder, a self-attention model and a decoder, and wherein: the encoder comprises a recurrent neural network which is configured to generate a sequence of hidden states based upon the sequence of meter data; the self-attention model is configured to generate a context vector based upon the sequence of hidden states; and the decoder is configured to generate the ranking scores for each of the plurality of transformers based upon the context vector.

8. The computer-implemented method of claim 7, wherein the decoder is configured to apply weights to the context vector and also comprises a function layer which is configured to generate the ranking scores for the plurality of transformers based upon the weighted context vector.

9. The computer-implemented method of any preceding claim, wherein the electricity meter comprises a multi-phase electricity meter or a three-phase electricity meter.

10. The computer-implemented method of any preceding claim, wherein a number of active phase(s) of the electricity meter is changeable in response to a variation in power load of the electricity meter.

11. The computer-implemented method of any preceding claim, wherein the electricity meter is a first electricity meter, and the method further comprises: generating a training dataset, wherein the training dataset comprises, for each of a first plurality of electricity meters: a sequence of meter data generated based upon measurements obtained from the respective electricity meter at a plurality of time points, and target label data indicative of a transformer which is one of the plurality of transformers; and training the machine learning model based upon the training dataset.

12. The computer-implemented method of claim 11, wherein generating the training dataset comprises:obtaining measurements from a second plurality of electricity meters at a plurality of time points, wherein the first electricity meters are included within the second plurality of electricity meters; obtaining information indicative of transformers which are associated with respective ones of the second plurality of electricity meters; and one or more of the following steps: calculating a distance between each of the second plurality of electricity meters and its associated transformer, and if the calculated distance related to a first one of the second plurality of electricity meters is longer than a distance threshold, excluding data associated with the first one of the second plurality of electricity meters from the training dataset; and determining whether any of the obtained measurements comprise anomalous measurements, and if the anomalous measurements were obtained from a second one of the second plurality of electricity meters, excluding data associated with the second one of the second plurality of electricity meters from the training dataset.

13. The computer-implemented method of claim 12, wherein generating the training dataset comprises generating the target label data based upon the information indicative of transformers which are associated with respective ones of the second plurality of electricity meters.

14. The computer-implemented method of claim 12 or 13, wherein generating the training dataset further comprises: obtaining geographical locations of the first plurality of electricity meters; and identifying a list of available transformers which are located in a neighbourhood of the first plurality of electricity meters, wherein the plurality of transformers are included within the list of available transformers.

15. The computer-implemented method of any one of claims 12 to 14, wherein generating the training dataset comprises: organising at least some of the first plurality of electricity meters which have highly correlated measurements into a sub-group, wherein the obtained information for the electricity meters within the sub-group is indicative of a first transformer;calculating a correlation coefficient of measurements between a second electricity meter of the first plurality of meters and an electricity meter of the sub-group, wherein the second electricity meter is not included within the sub-group; and if the correlation coefficient is higher than a correlation threshold, generating or modifying the target label data for the second electricity meter so as to indicate the first transformer.

16. The computer-implemented method of claim 15 as dependent from claim 14, wherein the obtained information indicates that a second transformer is associated with the second electricity meter, and generating the training dataset further comprises: determining whether the second transformer is within the list of available transformers; wherein calculating the correlation coefficient of measurements between the second electricity meter and an electricity meter of the sub-group is performed in response to determining that the second transformer is not within the list of available transformers.

17. The computer-implemented method of claim 15 or 16, wherein the correlation threshold is an average value of the correlation coefficients between the electricity meters of the sub-group.

18. The computer-implemented method of any one of claims 11 to 17, wherein the first plurality of electricity meters are within a geographical region, and the method further comprises: determining whether a location of the first electricity meter is within the geographical region, wherein providing the sequence of meter data to the machine learning model is performed if but only if the location of the first electricity meter is within the geographical region.

19. A system for performing meter-to-transformer mapping in an electrical distribution network, comprising: a processor; and a memory coupled to the processor storing instructions, which when executed by the processor, cause the processor to perform steps comprising:obtaining measurements from an electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; providing the sequence of meter data to a machine learning model; receiving from the machine learning model ranking scores for each of a plurality of transformers; and selecting one of the plurality of transformers which has the highest ranking score within the plurality of transformers for mapping with the electricity meter.

20. A method of training a machine learning model for performing meter-to- transformer mapping in an electrical distribution network, comprising: generating a training dataset, wherein the training dataset comprises, for each of a first plurality of electricity meters: a sequence of meter data generated based upon measurements obtained from the respective electricity meter at a plurality of time points, and target label data indicative of a transformer which is one of a plurality of transformers located in a neighbourhood of the first plurality of electricity meters, and wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; and training the machine learning model based upon the training dataset.

21. A method of obtaining a training dataset for training a machine learning model for performing meter-to-transformer mapping in an electrical distribution network, comprising, for each of a first plurality of electricity meters: obtaining measurements from the respective electricity meter at a plurality of time points; generating a sequence of meter data based upon the measurements, wherein the sequence of meter data is arranged in time order based upon the plurality of time points, and wherein the meter data within each entry of the sequence is generated based upon the measurements obtained at a respective one of the time points; andobtaining target label data indicative of a transformer which is one of a plurality of transformers located in a of the first plurality of electricity meters; wherein the training dataset comprises the sequence of meter data and the target label data of each of the first plurality of electricity meters.

Images