Intelligent electric energy meter local abnormal point detection method and system

Abstract

The application discloses a kind of intelligent electric energy meter local anomaly point detection method and system, it is related to the field of anomaly detection, it is by being labeled as normal in training stage, using historical power consumption data, define window length and sliding step, extract a series of normal power consumption data sliding window sequence.Succeedingly, based on these normal sliding window sequence training LSTM self-encoder, enable it to deep learning and internalize normal power consumption mode time sequence characteristics and inherent law.In detection stage, the same way of sliding window sampling is carried out to the data to be detected, and is input into the trained self-encoder.By calculating reconstruction error, this method can accurately identify the local anomaly point that significantly deviates from normal mode.This method overcomes the defect that traditional fixed threshold or global statistical method cannot capture subtle local changes, thereby significantly improving power supply reliability and operating efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of anomaly detection, and more particularly, to a smart meter local anomaly point detection method and system. BACKGROUND

[0002] As a core component of modern power systems, smart meters can collect massive amounts of user power consumption data in real time and accurately. These data are not only the basis for electricity settlement, but also the key information source for power companies to conduct load forecasting, equipment health monitoring, and identify abnormal behavior. However, in these vast data streams, there are some local anomalies that are short in duration, may not be extreme in amplitude, but are crucial. These local anomalies, such as a user's power consumption suddenly and significantly lower than their recent similar period during the summer air conditioning peak period, may be an early signal of electricity theft, equipment failure, or metering error. If these local anomalies can be identified in a timely and accurate manner, it will greatly improve the efficiency of the power system, reduce economic losses, and ensure the reliability of power supply. Therefore, it is of great practical value to build an efficient and accurate smart meter local anomaly point detection scheme.

[0003] However, existing anomaly detection methods face many challenges when dealing with smart meter data. Traditional methods based on fixed thresholds or global statistical distribution often fail to effectively capture these local anomalies. On the one hand, power consumption data has significant periodicity and trendiness, and the fluctuation amplitude of these global patterns is much larger than that of local anomalies, making local anomalies appear insignificant in a global perspective and easily submerged in normal fluctuations, forming a so-called global pattern masking effect. On the other hand, users' normal power consumption patterns are not fixed, and there are significant differences between different users. A fixed or global normal standard is difficult to adapt to all users and all time periods, making it difficult to judge local anomalies based on this standard. These limitations make it difficult for traditional methods to achieve fine identification of local anomalies when faced with complex and variable power consumption data, missing the opportunity to discover potential problems early.

[0004] In order to overcome the above challenges and achieve accurate identification of local anomalies, an intelligent detection method is needed that can fully utilize the time series characteristics of power consumption data and adaptively learn normal patterns. SUMMARY

[0005] In view of the above limitations of existing methods, according to an aspect of the present application, a smart meter local anomaly point detection method is provided, which includes a training phase and a detection phase.

[0006] The training phase includes:

[0007] Obtaining historical power consumption data labeled as normal;

[0008] define a window length and a sliding step, and extract a sequence of normal power consumption data sliding windows from the historical power consumption data labeled as normal;

[0009] train an LSTM autoencoder model based on the sequence of normal power consumption data sliding windows to obtain a trained LSTM autoencoder model;

[0010] a detection stage, comprising:

[0011] obtain to-be-detected power consumption data;

[0012] slide window sample the to-be-detected power consumption data to obtain a sequence of detection power consumption data sliding windows with the window length and the sliding step;

[0013] input the sequence of detection power consumption data sliding windows into the trained LSTM autoencoder model to obtain a reconstruction error sequence;

[0014] label a time point of abnormality occurrence based on the reconstruction error sequence.

[0015] According to another aspect of the present application, there is provided an intelligent electric energy meter local anomaly point detection system, comprising:

[0016] a training module, comprising:

[0017] a normal historical power consumption data obtaining unit configured to obtain historical power consumption data labeled as normal;

[0018] a normal power consumption data extracting unit configured to define a window length and a sliding step, and extract a sequence of normal power consumption data sliding windows from the historical power consumption data labeled as normal;

[0019] an LSTM autoencoder training unit configured to train an LSTM autoencoder model based on the sequence of normal power consumption data sliding windows to obtain a trained LSTM autoencoder model;

[0020] a detection module, comprising:

[0021] a to-be-detected power consumption data obtaining unit configured to obtain to-be-detected power consumption data;

[0022] a detection power consumption data sliding sampling unit configured to slide window sample the to-be-detected power consumption data to obtain a sequence of detection power consumption data sliding windows with the window length and the sliding step;

[0023] a reconstruction error generating unit configured to input the sequence of detection power consumption data sliding windows into the trained LSTM autoencoder model to obtain a reconstruction error sequence;

[0024] An abnormal time point marking unit is configured to mark a time point at which an anomaly occurs based on the reconstruction error sequence.

[0025] Compared with the prior art, the intelligent electric energy meter local anomaly point detection method and system provided by the application can define the window length and sliding step by using the historical power consumption data labeled as normal in the training stage, and extract a series of normal power consumption data sliding window sequences. These sliding windows cleverly capture the local timing context information of the power consumption data, effectively solve the global pattern masking effect and the dynamic and heterogeneous problems of the normal mode in the background technology, and ensure that the local anomaly will not be overwhelmed by the macro fluctuations. Subsequently, based on these normal sliding window sequences, the LSTM autoencoder is trained, so that it can deeply learn and internalize the timing features and internal rules of the normal power consumption mode, rather than relying on fixed global standards. In the detection stage, the same way of sliding window sampling is performed on the data to be detected, and the trained autoencoder is input. By calculating the reconstruction error, this method can accurately identify the local anomaly points that deviate significantly from the normal mode. This method overcomes the defect that the traditional fixed threshold or global statistical method cannot capture subtle local changes, realizes the early and accurate discovery of key signals such as electricity stealing and equipment failure, and thus significantly improves the power supply reliability and operation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0026] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, and do not constitute a limitation of the present application. In the drawings, like reference numerals refer to like parts or steps throughout.

[0027] Figure 1 A flowchart of the training stage in the intelligent electric energy meter local anomaly point detection method according to the embodiments of the present application.

[0028] Figure 2 A flowchart of step S130 in the intelligent electric energy meter local anomaly point detection method according to the embodiments of the present application.

[0029] Figure 3 A flowchart of the detection stage in the intelligent electric energy meter local anomaly point detection method according to the embodiments of the present application.

[0030] Figure 4 A data flow diagram of the detection stage in the intelligent electric energy meter local anomaly point detection method according to the embodiments of the present application.

[0031] Figure 5 A flowchart of step S240 in the intelligent electric energy meter local anomaly point detection method according to the embodiments of the present application.

[0032] Figure 6 A block diagram of a local anomaly point detection system of a smart electric energy meter according to an embodiment of the present application. DETAILED DESCRIPTION

[0033] Embodiments of the present disclosure will be described in more detail with reference to the drawings. While certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood. It should be understood that the drawings of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0034] To solve the problems in the background art, the present application provides a local anomaly point detection method for a smart electric energy meter. Figure 1 A flowchart of a training phase of a local anomaly point detection method for a smart electric energy meter according to an embodiment of the present application. As shown in Figure 1 The training phase of the local anomaly point detection method for a smart electric energy meter according to an embodiment of the present application includes: S110, obtaining historical power consumption data labeled as normal; S120, defining a window length and a sliding step, and extracting a sequence of normal power consumption data sliding windows from the historical power consumption data labeled as normal; and S130, training an LSTM autoencoder model based on the sequence of normal power consumption data sliding windows to obtain a trained LSTM autoencoder model.

[0035] In step S110, the historical power consumption data labeled as normal is obtained. It should be understood that the LSTM autoencoder model used in the present scheme aims to learn and internalize the inherent rules and time sequence characteristics of normal power consumption patterns. If abnormal data is mixed in the training data, the model will not be able to accurately distinguish between normal and abnormal, resulting in a decline in its ability to recognize abnormalities. By training only using historical data labeled as normal, the model can build a pure normal pattern representation, so that in the detection phase, any data point that deviates significantly from this normal pattern can be effectively identified as abnormal, thereby overcoming the challenges posed by the dynamic nature and heterogeneity of normal patterns in the background art, ensuring the accuracy and robustness of anomaly detection.

[0036] Specifically, one implementation of step S110 is as follows: first, obtain the historical power consumption data of a specific user or user group from the smart meter data acquisition system or the power data center. These data are stored in the form of time series, containing time stamps and corresponding power consumption, for example, power consumption every 15 minutes, every hour or every day. Second, perform preliminary data cleaning on these raw historical power consumption data, including handling missing values, such as filling in by interpolation or forward filling, removing obvious data acquisition errors or outliers, such as negative power consumption or extreme values far exceeding the physical limit.

[0037] Subsequently, the normal labeling process is performed. This process can combine various methods. One common and reliable method is to cooperate with the operation and maintenance personnel or experts in the field of the power company, and according to their experience and historical event records, such as no electricity stealing report, no equipment failure record, no metering abnormal complaint, etc., to screen out the power consumption data confirmed to have no abnormal occurrence in a specific time period. For example, the continuous months of power consumption data of the user in the past year without any abnormal event report can be selected as the initial normal data set. Another auxiliary method is to use some mature but insensitive to local anomalies statistical methods or rules to pre-screen the preliminary cleaned data, and exclude the obvious global anomalies, so as to obtain a relatively clean data set, and then the final confirmation is made by the experts. For example, a preliminary threshold based on historical mean and standard deviation can be set, and the data points exceeding the threshold range are marked as potential anomalies and excluded, so as to ensure that the data set input to the LSTM autoencoder training represents the true normal power consumption behavior as much as possible.

[0038] In step S120, the window length and the sliding step are defined, and the sequence of normal power consumption data sliding windows is extracted from the historical power consumption data labeled as normal. It can be understood that, as a typical time series data, the judgment of local anomaly of power consumption data strongly depends on its context. A single isolated data point is difficult to be accurately judged as abnormal, because it may be only a part of normal fluctuation. Therefore, in the technical solution of the present application, by defining the window length, the continuous power consumption data can be divided into segments with a certain time span, each segment contains the data point and its adjacent data before and after, thereby providing the necessary context information for judging whether the power consumption pattern in this time period is normal, effectively solving the problem of dependence of the definition of local anomaly on context as pointed out in the background art. At the same time, the introduction of sliding step makes the windows overlap, which not only increases the diversity of training data, but also enables the model to capture the subtle changes and trends in the data sequence in more detail, avoiding the risk of missing potential anomalies due to improper window division, and further enhancing the learning ability of the model to the dynamics and heterogeneity of normal patterns.

[0039] In particular, one implementation of step S120 is as follows: First, the window length and sliding step need to be determined according to the actual application scenario and the sampling frequency of the power consumption data. For example, if the smart power meter collects power consumption data at an hourly unit, and experience shows that local anomalies such as electricity theft or equipment transient failure usually last for several hours, the window length can be set to 8 hours, that is, each sliding window contains 8 consecutive power consumption data points. This length is sufficient to cover the typical duration of local anomalies and provides sufficient contextual information. The sliding step can be set to 1 hour, that is, every time a new window is generated, the data sequence is shifted forward by 1 hour. This setting makes there is a 7-hour overlap between adjacent windows, ensuring that any subtle changes in the data sequence can be captured by multiple windows, thereby providing more abundant learning samples for the LSTM autoencoder.

[0040] After determining the window length and sliding step, a sequence of normal power consumption sliding windows is extracted from the obtained historical power consumption data labeled as normal, for example, a time series containing continuous monthly hourly power consumption: , ,..., ]. The specific operation is: starting from the beginning of the time series, the first data segment determined by the window length is intercepted as the first sliding window, for example, the first window is[ , ,..., ]. Then, according to the preset sliding step, for example, 1 hour, the interception point is moved forward, and the next data segment is intercepted as the second sliding window, that is, , ,..., ]. Repeat this process until the end of the time series, that is, stop when there is less than one window length of data left. Finally, arrange all the intercepted sliding windows in the order of their generation to form a sequence of normal power consumption sliding windows.

[0041] In step S130, an LSTM autoencoder model is trained based on the sequence of the normal electricity consumption data sliding window to obtain a trained LSTM autoencoder model. It is worth noting that the LSTM autoencoder, as a deep learning model, excels at processing time-series data and learning its inherent temporal dependencies and patterns. Therefore, this application trains the model by inputting the sequence of the normal electricity consumption data sliding window, enabling the autoencoder to learn how to efficiently encode and decode these normal patterns. That is, by training using only historical electricity consumption data labeled as normal, the autoencoder can construct an internal representation highly sensitive to normal patterns. When the model encounters data significantly different from the normal patterns seen during training—i.e., anomalous data—during the detection phase, it will struggle to accurately reconstruct the data, resulting in a large reconstruction error. This difference in reconstruction error is the core basis for identifying local anomalies, enabling the model to effectively distinguish between normal fluctuations and genuine abnormal behavior, overcoming the challenge of traditional methods' inability to adapt to complex electricity consumption patterns.

[0042] Specifically, in one possible implementation of this application, Figure 2 This is a flowchart of step S130 in the smart energy meter local anomaly detection method according to an embodiment of this application. Figure 2 As shown, step S130, training an LSTM autoencoder model based on the sequence of the normal electricity consumption data sliding window to obtain a trained LSTM autoencoder model, includes: S131, inputting the sequence of the normal electricity consumption data sliding window into the encoder of the LSTM autoencoder model to obtain a sequence of the temporal context hidden encoding vector of the normal electricity consumption data; S132, inputting the sequence of the temporal context hidden encoding vector of the normal electricity consumption data into the decoder of the LSTM autoencoder model to obtain a sequence of the reconstructed electricity consumption data sliding window; S133, calculating a loss function value based on the sequence of the normal electricity consumption data sliding window and the sequence of the reconstructed electricity consumption data sliding window; S134, training the LSTM autoencoder model based on the loss function value and through backpropagation of gradient descent.

[0043] More specifically, in an implementation of the present application, the encoder of the LSTM autoencoder model is one or more LSTM layers, and the decoder of the LSTM autoencoder model is one or more LSTM layers. In particular, LSTM layers selectively memorize, update and output information through cell state and three gating mechanisms, namely forget gate, input gate and output gate. It is worth mentioning that both the encoder and the decoder of the LSTM autoencoder model employ one or more LSTM layers, which is based on the inherent advantages of LSTM in processing time series data and the powerful ability of deep learning models in feature learning. As a special recurrent neural network unit, LSTM layers can effectively solve the problem of gradient vanishing or gradient explosion in traditional recurrent neural networks when processing long sequences, so that they can capture long-term dependencies and complex temporal patterns in electricity consumption data. Specifically, the encoder is designed as one or more LSTM layers, which can gradually extract and compress high-dimensional, abstract temporal context hidden encoding vectors from the sequence of input normal electricity consumption data sliding windows. The multi-layer LSTM structure can enhance the representation ability of the model, so that it can learn electricity consumption pattern features at different levels, for example, the first layer may learn short-term intra-day fluctuations, while deeper layers may capture weekly, monthly or even seasonal electricity consumption rules, thereby more comprehensively and finely understanding the complexity of normal electricity consumption behavior. Similarly, the decoder employs one or more LSTM layers, which can gradually reconstruct the original electricity consumption data sliding window sequence based on the hidden vectors extracted by the encoder. This symmetrical structure enables the model to better understand the data generation mechanism during the learning process.

[0044] It should be understood that the encoder is the core component of the autoencoder model, which aims to learn and extract low-dimensional and efficient feature representations of the input data. For smart meter data, this means that the encoder needs to distill the inherent rules and structures from the normal electricity consumption patterns containing temporal context. Therefore, the present application encodes the data to remove redundant information in the data while retaining features essential for identifying normal patterns to obtain normal electricity consumption data temporal context hidden encoding vectors. This hidden encoding vector is the model's deep understanding of normal electricity consumption behavior, which provides a basis for the subsequent decoder to reconstruct the original data and ensures that the model can sensitively capture any abnormal behavior deviating from this normal pattern during the detection phase.

[0045] In particular, one implementation of step S131 is as follows: the encoder of the LSTM autoencoder model can be designed as one or more LSTM layers. Taking a two-layer LSTM encoder as an example, its architecture is as follows: the first-layer LSTM receives the normal electricity consumption data points of each sliding window as input. For example, if each normal electricity consumption data point is a single-variable value, the input feature dimension of the first-layer LSTM is 1. This layer can be configured with, for example, 64 LSTM units, i.e., the dimension of the hidden state, and set to return sequence output so as to pass the hidden state of each time step to the next layer. The second-layer LSTM receives the sequence output of the first-layer LSTM as input, and can be configured with, for example, 32 LSTM units, but is set to return only the final hidden state. This design enables the encoder to output a fixed-dimension vector after processing the entire sliding window, which is the normal electricity consumption data time-series context implicit encoding vector of the sliding window.

[0046] The encoding process is performed independently for each normal electricity consumption data sliding window extracted. For each sliding window in the sequence, e.g., , ,..., , the encoder processes the data points therein in time sequence one by one. When is input into the first-layer LSTM, it updates its internal hidden state and cell state; then is input, the state continues to be updated, until is input. The output sequence of the first-layer LSTM is then input into the second-layer LSTM, which is also processed point by point. When the second-layer LSTM finishes processing the last data point in the window, its final hidden state, a 32-dimensional vector, is extracted as the normal electricity consumption data time-series context implicit encoding vector of the sliding window.

[0047] This encoding process traverses all the normal electricity consumption data sliding windows in the sequence of normal electricity consumption data sliding windows. After processing each sliding window, a corresponding normal electricity consumption data time-series context implicit encoding vector is obtained. Finally, these encoding vectors are arranged in the order of their corresponding sliding windows to form a sequence of normal electricity consumption data time-series context implicit encoding vectors. It is worth noting that the weights and bias parameters contained in the LSTM layers in the encoder are randomly initialized at the beginning of training, and are iteratively optimized through the gradient descent and backpropagation algorithm in step S134 to minimize the reconstruction error.

[0048] It can be understood that the design goal of the autoencoder model is to learn an effective, low-dimensional representation of the input data, and to verify the effectiveness of this representation by reconstructing the original data. The role of the decoder is to map the implicit encoding vector extracted by the encoder from the normal power consumption data, which compresses the temporal context information, back to the original data space. Through this reconstruction process, the model is forced to learn and master the internal rules and structure of the normal power consumption pattern. Only when the model can accurately reconstruct the normal power consumption data, it indicates that it has successfully learned the characteristics of these normal patterns.

[0049] Specifically, one implementation of step S132 is as follows: the decoder of the LSTM autoencoder model is also designed as one or more LSTM layers, which has a similar structure to the encoder. Continuing the example of the encoder in step S131, if the encoder outputs a 32-dimensional implicit encoding vector of the temporal context of the normal power consumption data, the decoder will receive this vector as its input. A typical two-layer LSTM decoder architecture is as follows: the first layer LSTM of the decoder receives the 32-dimensional implicit encoding vector from the encoder. This vector can be input to the decoder LSTM layer as the initial hidden state, or repeated at each time step. This layer can be configured with, for example, 32 LSTM units, and set to return a sequence output. The second layer LSTM receives the sequence output of the first layer LSTM as input, and can be configured with, for example, 64 LSTM units, also set to return a sequence output. After the output of each time step of the second layer LSTM, a fully connected layer is connected, which converts the hidden state of the LSTM to an output matching the dimension of the original power consumption data points, for example, a single variable power consumption value.

[0050] The decoding process is carried out independently for each normal power consumption data temporal context implicit encoding vector. For each implicit encoding vector, the decoder will attempt to generate a sequence of the same length as the original sliding window. For example, if the original sliding window length is 8 hours, the decoder will generate 8 consecutive power consumption data points. The decoder, through its internal LSTM layer, predicts the power consumption value at the current time step based on the output and internal state of the previous time step at each time step. This process continues until a sequence of the same length as the original sliding window is generated. Finally, all the sliding windows reconstructed by the decoder are arranged in the order of their corresponding implicit encoding vectors to form a sequence of reconstructed power consumption data sliding windows. It is worth mentioning that, similar to the encoder, it is randomly initialized at the beginning of training, and iteratively optimized through gradient descent and backpropagation algorithm in step S134 to minimize the reconstruction error.

[0051] In particular, here, when defining the window length and the sliding step length in the training process, usually the sliding step length is set to be smaller than the window length, so that when the sequence of the normal power consumption data sliding window is time series encoded by the LSTM auto-encoder model, the redundant time series precursor noise distribution is accumulated due to the time series repetition of the time series window based on the sliding step length, and further, the time series correlation expansion is embodied in the near-far bidirectional time series context association, thereby interfering with the training of the LSTM auto-encoder model.

[0052] Based on this, preferably, in another implementable manner of the present application, step S132, inputting the sequence of the normal power consumption data time series context implicit encoding vectors into the decoder of the LSTM auto-encoder model to obtain the sequence of the reconstructed power consumption data sliding window, comprises:

[0053] Firstly, based on the window length and the sliding step length, the time series redundant precursor probability is defined, that is:

[0054]

[0055] wherein, is the window length, is the sliding step length, is the time series redundant precursor probability;

[0056] Then, the sequence of the normal power consumption data time series context implicit encoding vectors is locally sparsified based on the time series redundant precursor probability to obtain a local sparsification regulation factor, that is, by introducing the local sparsification of the time series redundant precursor probability as the core through the norm representation of the discrete weight:

[0057]

[0058] wherein, is the vector after merging the sequence of the normal power consumption data time series context implicit encoding vectors, is the power of , and is the power of , and is the power of , and is the local sparsification regulation factor, that is, the local sparsification identifier of the norm representation of the feature set, the time series redundant precursor probability is used for redundant synergy elimination;

[0059] Then, the time redundancy predecessor probability and the local sparsification regulatory factor act on the feature value of each normal power consumption data time context implicit coding vector in the sequence of normal power consumption data time context implicit coding phase difference coordination vectors in a negotiated resolution form to obtain the sequence of normal power consumption data time context implicit coding phase difference coordination vectors, that is,

[0060]

[0061] wherein, is each feature value in the sequence of normal power consumption data time context implicit coding phase difference coordination vectors, is the feature value of each vector in the sequence of normal power consumption data time context implicit coding phase difference coordination vectors, so that the redundancy synergy resolution and the time redundancy predecessor act on each feature value in a negotiated resolution form, in particular, here is the coordinated is split according to the form before merging to obtain the sequence of normal power consumption data time context implicit coding phase difference coordination vectors;

[0062] Finally, the sequence of normal power consumption data time context implicit coding phase difference coordination vectors is input into the decoder of the LSTM autoencoder model to obtain the sequence of reconstructed power consumption data sliding windows, so that the temporal interference in the training process can be reduced in a way of flexibly eliminating the redundancy noise through negotiation. That is, by interacting the local sparsity identifier and the time redundancy predecessor noise through phase coupling, a connection is established in a negotiated resolution form, that is, by accurately acting on the local redundancy disturbance source of the feature value granularity through global phase interaction, the phase difference coordination process of the cooperative redundancy is simulated, so that the resolution strategy can effectively suppress the redundancy noise and flexibly act on the stability requirement of global features, thereby improving the training effect of the LSTM autoencoder model. In particular, the decoding process here is the same as the above embodiment.

[0063] That is, in the training phase of the smart power meter local anomaly point detection method, calculating the loss function value is a key link to evaluate the learning effect of the LSTM autoencoder model. In order to quantify the accuracy of the model in reconstructing normal power consumption data, that is, the difference between the original input and the model output. The present application obtains an explicit optimization target by calculating this difference, so as to continuously adjust its internal parameters in the subsequent training process in order to minimize this difference, so that the model can more accurately learn and represent the normal power consumption mode.

[0064] More specifically, in one implementation of the present application, the step S133 of calculating the loss function value based on the sequence of normal power consumption data sliding windows and the sequence of reconstructed power consumption data sliding windows comprises: calculating the mean square error of the sequence of normal power consumption data sliding windows and the sequence of reconstructed power consumption data sliding windows as the loss function value.

[0065] The implementation is as follows: first, from the sequence of normal power consumption data sliding windows, a current original sliding window is taken out, for example, a time sequence containing 8 consecutive hourly power consumption data points: , , , , , , , ]. At the same time, from the obtained sequence of reconstructed power consumption data sliding windows, a reconstructed sliding window corresponding to the original window is taken out, for example: , , , , , , , ].

[0066] Next, for each corresponding data point in the two sequences, the square of the difference value is calculated. For example, for the first data point, the square of the difference ( - ) is calculated; for the second data point, the square of the difference ( - ) is calculated, and so on until the last data point ( - )^2. Then, the square difference values are accumulated and divided by the length of the sliding window, which is 8 in this example, to obtain the mean square error of the single sliding window. For example, if the specific power consumption data of the original window is [10, 12, 11, 15, 14, 13, 10, 9] and the corresponding window data reconstructed by the model is [10.2, 11.9, 11.1, 15.3, 13.8, 13.1, 9.8, 9.2], the square differences of each point are: for the first data point (10-10.2)^2=0.04, for the second data point (12-11.9)^2=0.01... and so on until the eighth data point (9-9.2)^2=0.04. The accumulated sum is 0.28. The mean square error of the window is 0.28 / 8=0.035.

[0067] This MSE value represents the accuracy of the model in reconstructing the current sliding window. In actual training, data is processed in batches, i.e. the MSE of all sliding windows in a batch is calculated at once, and then these MSEs are averaged to obtain the loss function value of the current batch.

[0068] Correspondingly, the internal parameters, weights and biases of the autoencoder model are random or preset in the initial stage, and cannot accurately reconstruct normal power consumption data. The calculated loss function value quantifies the gap between the current reconstruction ability of the model and the real normal mode. In order to narrow this gap, the model needs a mechanism to adjust its internal parameters. Gradient descent and backpropagation are such a mechanism, which can efficiently calculate the influence direction and size of each parameter on the loss function, and iteratively update the parameters accordingly, so that the reconstruction error of the model gradually decreases after each iteration. Through continuous training, the model can finally learn and internalize the complex time series pattern of normal power consumption data, and accurately distinguish between normal and abnormal in the detection stage.

[0069] Specifically, one implementation of step S134 is as follows: first, before the start of training, the weight and bias parameters of the LSTM autoencoder model (including all LSTM layers and possible fully connected layers in the encoder and decoder) are randomly initialized. After completing a forward propagation in steps S131 and S132, and calculating the loss function value of the current batch of data by step S133, the training enters the backpropagation phase.

[0070] The backpropagation algorithm starts from the output layer of the model, i.e. the reconstructed power consumption data sliding window reconstructed by the decoder, and calculates the gradient of the loss with respect to each parameter in the model according to the loss function value. This process uses the chain rule to pass error signals from the output layer to the previous layers, calculating the contribution of the weights and biases of each layer, each neuron, or LSTM unit to the total loss. For example, if the loss function is MSE, the error signal will be calculated according to the difference between the reconstructed value and the real value, and propagated along the reverse path of the network.

[0071] After obtaining the gradients of all parameters, the gradient descent optimizer updates the weights and biases of the model according to these gradients. The update rule is: new parameter = old parameter - learning rate * gradient. Among them, the learning rate is a key hyperparameter that determines the step size of each parameter update. The setting of the learning rate is crucial, for example, it can be set to 0.001. If the learning rate is too large, the model may not converge, and the loss value will fluctuate or even diverge; if the learning rate is too small, the model will converge very slowly. In practical applications, in addition to the basic gradient descent, adaptive learning rate optimization algorithms such as Adam and RMSprop are more commonly used. They can dynamically adjust the learning rate according to the historical information of the gradient, thereby accelerating the convergence and improving the training stability. This forward propagation, loss calculation, back propagation, and parameter update cycle will be repeated until the loss function value converges below the preset threshold, or the maximum training round is reached, for example, 100 epochs, or the performance on the validation set no longer improves, at this time the trained LSTM autoencoder model is obtained.

[0072] After training is completed, enter the detection phase. That is, after training the LSTM autoencoder model using the smart meter local anomaly point detection algorithm, the trained LSTM autoencoder model is used in the actual detection phase.

[0073] Figure 3 The flow chart of the detection phase in the smart meter local anomaly point detection method according to the embodiments of the present application. Figure 4 The data flow diagram of the detection phase in the smart meter local anomaly point detection method according to the embodiments of the present application. As shown in Figure 3 and Figure 4 The detection phase of the smart meter local anomaly point detection method according to the embodiments of the present application includes: S210, obtaining the to-be-detected power consumption data; S220, performing sliding window sampling on the to-be-detected power consumption data to obtain a sequence of detection power consumption data sliding windows with the window length and the sliding step; S230, inputting the sequence of detection power consumption data sliding windows into the trained LSTM autoencoder model to obtain a reconstruction error sequence; S240, based on the reconstruction error sequence, marking the time point of abnormal occurrence.

[0074] In step S210, the to-be-detected power consumption data is acquired. It should be understood that the training phase has enabled the LSTM autoencoder model to learn and master the internal rules of the normal power consumption mode. In order to apply this learning ability to the actual scene and timely discover potential local anomalies such as electricity theft, equipment failure or metering error, the model needs to continuously receive the latest and never-before-seen power consumption data from the smart meter. Only by acquiring these real-time or near real-time data can they be input into the trained model for analysis, and whether there is an abnormal behavior deviating from the normal mode can be judged according to the reconstruction error output by the model. Therefore, this step is the basis for realizing continuous monitoring and anomaly warning of power consumption data.

[0075] Specifically, one implementation of step S210 is as follows: acquiring the to-be-detected power consumption data involves continuously receiving power consumption data streams from smart meter data acquisition devices or power data aggregation equipment. These data are collected at a preset sampling frequency, for example, recording power consumption values every hour or every 15 minutes, with accurate timestamps. For example, the power consumption data of a certain user on a specific date, such as October 26, 2023, may continuously flow in the form of a time series, such as: [2023-10-26 00:00:00, 1.2kWh], [2023-10-26 01:00:00, 1.5 kWh], [2023-10-26 02:00:00, 1.3kWh], etc.

[0076] After the data is acquired, necessary preprocessing is required to ensure that its format and characteristics are consistent with the data in the training phase. This includes: first, basic data integrity checks, such as verifying the continuity of timestamps and the validity of power consumption values, such as non-negative values. For a small number of missing values or obvious error values that may occur, interpolation, forward filling or direct rejection strategies can be used for processing, but care should be taken to avoid introducing new abnormal information. After these processes, the to-be-detected power consumption data is obtained.

[0077] In step S220, the to-be-detected power consumption data is sampled with the window length and the sliding step to obtain a sequence of detection power consumption data sliding windows. Accordingly, in order for the model to accurately assess whether the new power consumption data is normal, the to-be-detected data needs to be presented to the model in the same format and context as the training data. A single isolated power consumption data point cannot provide enough time series information to determine whether it is abnormal, because the definition of local anomaly depends on its context environment. By using the same window length and sliding step as in the training phase for sliding window sampling, it can be ensured that the input data received by the model has the same time series structure and context information, so that it can use the normal mode it has learned to effectively compare and reconstruct, and then identify local anomalies that deviate from normal behavior.

[0078] Specifically, one implementation of step S220 is as follows: It is worth mentioning that the window length and the sliding step have been determined and fixed in the training stage. For example, if the window length is set to 8 hours and the sliding step is set to 1 hour in the training stage, then these two parameters are strictly followed in the detection stage.

[0079] First, starting from the beginning of the time series of the to-be-detected power consumption data, the first data segment with a length equal to the window length is intercepted. For example, if the to-be-detected data is a continuous hourly power consumption sequence [ , , ,..., ], and the window length is 8, then the first detection power consumption sliding window will be [ , , , , , , , ].

[0080] Next, the interception point is moved forward by the preset sliding step. For example, if the sliding step is 1, then the interception point is moved forward by one time step, and the second detection power consumption sliding window is intercepted: [ , , , , , , , ]. This process continues until the end of the to-be-detected power consumption data sequence, i.e., when there is not enough data for a complete window length.

[0081] Finally, all the sliding windows obtained in this way are arranged in the time order in the original to-be-detected data to form a sequence of detection power consumption sliding windows.

[0082] In step S230, the sequence of detection electricity data sliding windows is input into the trained LSTM autoencoder model to obtain a sequence of reconstruction errors. That is, the trained LSTM autoencoder model has learned and internalized the complex temporal features of normal electricity usage patterns. When new, to-be-detected electricity data is input into this model, if the data conforms to the normal patterns learned by the model, the model can accurately reconstruct it, resulting in a lower reconstruction error. Conversely, if the to-be-detected data contains abnormal behavior, i.e., deviates from the normal patterns learned by the model, the model will have difficulty accurately reconstructing it, resulting in a significantly higher reconstruction error. Therefore, this step provides a key quantitative basis for subsequent anomaly judgment by quantifying the difference between the input data and the normal representation inside the model (i.e., the reconstruction error), enabling the model to adaptively identify local outliers that deviate from normal behavior and overcoming the challenge of traditional methods that are difficult to adapt to dynamic electricity usage patterns.

[0083] Specifically, one implementation of step S230 is as follows: for each detection electricity data sliding window in the sequence, it is first input into the encoder part of the trained LSTM autoencoder model. The role of the encoder is to compress this original, high-dimensional electricity data sliding window into a low-dimensional temporal context hidden encoding vector. This vector is a compact representation of the original window, capturing its core temporal features. For example, an 8-hour electricity data window may be compressed into a 32-dimensional vector after being processed by the encoder.

[0084] Subsequently, the temporal context hidden encoding vector generated by the encoder is input into the decoder part of the trained LSTM autoencoder model. The task of the decoder is to attempt to expand and reconstruct the hidden encoding vector back into a time series with the same length and structure as the original input window, i.e., to reconstruct the electricity data sliding window. For example, the decoder attempts to restore an 8-hour electricity data sequence from the 32-dimensional hidden vector.

[0085] Finally, for each original detection electricity data sliding window and its corresponding reconstructed electricity data sliding window, the mean squared error between them is calculated according to the method described in step S133. Specifically, the difference between each corresponding data point in the two windows is squared, then summed and divided by the window length. The above process is performed for each window in the sequence of detection electricity data sliding windows. Each window processed results in a corresponding reconstruction error value. These reconstruction error values are arranged in the time order of their corresponding sliding windows, ultimately forming a sequence of reconstruction errors.

[0086] In step S240, a time point of abnormality occurrence is marked based on the reconstruction error sequence. It should be understood that the step S230 has calculated the reconstruction error for each sliding window of the power consumption data to be detected, and the error values quantify the deviation between the current power consumption mode and the normal mode learned by the model. However, only the error values cannot directly indicate the occurrence of abnormality. In order to convert these quantified errors into operational abnormality alarms and determine the time of abnormality occurrence, a judgment standard is needed to distinguish between normal fluctuations and real abnormalities. For this purpose, the present application can achieve accurate identification and timely warning of local abnormalities by determining the time point of abnormality occurrence, thereby providing a basis for subsequent intervention measures.

[0087] Specifically, in an implementable manner of the present application, Figure 5 The flowchart of step S240 in the local abnormality point detection method of the smart power meter according to the embodiment of the present application is shown in FIG. 4. As shown in FIG. 4, in step S240, a time point of abnormality occurrence is marked based on the reconstruction error sequence, including: S241, comparing each reconstruction error in the reconstruction error sequence with a preset reconstruction error threshold; S242, if the reconstruction error is greater than the preset reconstruction error threshold, marking that the corresponding detection sliding window of power consumption data contains a local abnormality; and S243, determining a time point of abnormality occurrence based on the corresponding detection sliding window of power consumption data. Figure 5

[0088] Specifically, an implementable manner of step S240 is as follows: first, S241 is executed. The setting of the preset reconstruction error threshold is determined after the model training is completed, using a part of normal power consumption data (validation set) that does not participate in the training. A common method is to calculate the reconstruction error distribution of these normal data on the model, and then select a statistical quantity as the threshold, for example, the mean value of the normal reconstruction error plus three times the standard deviation, or select a high percentile, such as the 99.5th percentile, to ensure that the reconstruction error of most normal data points is lower than the threshold. For example, if it is determined through statistical analysis that the threshold is 0.05. Then, each error value in the reconstruction error sequence will be compared with 0.05.

[0089] Then, S242 is executed. That is, when the reconstruction error of a certain sliding window, for example, 0.08, exceeds the preset threshold 0.05, the window is determined as an abnormal window. This marking process will traverse the entire reconstruction error sequence to identify all sliding windows determined as abnormal.

[0090] ​Finally, S243 is executed. Each detection electricity data sliding window marked as containing a local anomaly corresponds to a segment of continuous electricity data and its timestamp. For example, the window length is 8 hours, and the data is collected hourly. An anomaly sliding window may contain the electricity consumption data of 8 consecutive hours starting from a certain starting time point, such as 10:00 on October 26, 2023: [10:00, 11:00, 12:00, 13:00, 14:00, 15:00, 16:00, 17:00]. Specifically, in an implementation of the present application, based on the corresponding detection electricity data sliding window, the time point at which the anomaly occurs is determined, including: marking the timestamp of the center data point of the corresponding detection electricity data sliding window as the time point at which the anomaly occurs. Specifically, when a sliding window is determined to be abnormal, the abnormal behavior may occur at any time within the window. Simply selecting the starting or ending point of the window may not accurately reflect the core occurrence period of the anomaly. The center data point provides a balanced and representative estimate of the anomaly time within the window. It avoids biasing the anomaly time point to either end of the window, providing a relatively fair time reference for subsequent analysis or alarm, especially when further refinement is not possible or necessary to locate a specific data point within the window that contributes most to the anomaly. This is a practical and simple positioning strategy aimed at providing a representative time anchor point for the anomaly event. In order to determine the center data point of the sliding window, it needs to be calculated according to the length of the window. For a sliding window containing N data points, the center data point refers to the data point located in the middle position of the time series. If N is odd, the center point is clear; if N is even, such as 8 in this example, one of the two middle data points is selected, or more precisely, the data point closest in time to the midpoint of the window time span is selected. For example, for an 8-hour window from 10:00 to 17:00, the time midpoint is 13:30. The data point closest to this midpoint is 13:00, the 4th data point or 14:00, the 5th data point. In this implementation, one of them can be agreed upon, for example, the 4th data point, the data point with a timestamp of 13:00, is selected.

[0091] Finally, the timestamp of this center data point, for example, October 26, 2023, 13:00, is marked and output as the time point at which the anomaly represented by the anomaly sliding window occurs. This marked time point will be used for subsequent anomaly reporting, alarm triggering, or further analysis, thereby achieving accurate positioning of local anomalies in smart meter electricity data.

[0092] In summary, the intelligent electric energy meter local anomaly point detection method based on the embodiments of the present application is illustrated. In the training stage, the history electricity data labeled as normal is used to define the window length and the sliding step, and a series of normal electricity data sliding windows are extracted. These sliding windows cleverly capture the local timing context information of the electricity data, effectively solve the global pattern masking effect and the dynamic and heterogeneous problems of the normal mode in the background art, and ensure that the local anomaly will not be overwhelmed by the macro fluctuations. Then, based on the sequence of normal sliding windows, the LSTM autoencoder is trained, so that it can deeply learn and internalize the timing features and internal rules of the normal electricity mode, rather than relying on fixed global standards. In the detection stage, the same way of sliding window sampling is performed on the to-be-detected data, and the trained autoencoder is input. By calculating the reconstruction error, this method can accurately identify the local anomaly points that deviate significantly from the normal mode. This method overcomes the defect that the traditional fixed threshold or global statistical method cannot capture subtle local changes, realizes the early and accurate discovery of key signals such as electricity stealing and equipment failure, and thus significantly improves the power supply reliability and operation efficiency.

[0093] Figure 6 The figure is a block diagram of the intelligent electric energy meter local anomaly point detection system according to the embodiments of the present application. As shown in the figure, Figure 6 The intelligent electric energy meter local anomaly point detection system 100 according to the embodiments of the present application includes a training module 110, which includes a normal history electricity data acquisition unit 111 for acquiring history electricity data labeled as normal; a normal electricity data extraction unit 112 for defining the window length and the sliding step, and extracting a sequence of normal electricity data sliding windows from the history electricity data labeled as normal; and an LSTM autoencoder training unit 113 for training an LSTM autoencoder model based on the sequence of normal electricity data sliding windows to obtain a trained LSTM autoencoder model. The system also includes a detection module 120, which includes a to-be-detected electricity data acquisition unit 121 for acquiring to-be-detected electricity data; a detection electricity data sliding sampling unit 122 for sliding window sampling of the to-be-detected electricity data with the window length and the sliding step to obtain a sequence of detection electricity data sliding windows; a reconstruction error generation unit 123 for inputting the sequence of detection electricity data sliding windows into the trained LSTM autoencoder model to obtain a reconstruction error sequence; and an abnormal time point marking unit 124 for marking the time points of abnormal occurrence based on the reconstruction error sequence.

[0094] Here, those skilled in the art can understand that the specific operations of each step in the above intelligent electric energy meter local anomaly point detection system have been described in detail above with reference to the description of the intelligent electric energy meter local anomaly point detection method of Figures 1 to 5 Therefore, the repeated description will be omitted.

Claims

1. A method for detecting local abnormal points of a smart meter, characterized in that, The method comprises: a training phase and a detection phase; the training phase comprises: obtaining historical power consumption data labeled as normal; defining a window length and a sliding step, and extracting a sequence of normal power consumption data sliding windows from the historical power consumption data labeled as normal; training an LSTM autoencoder model based on the sequence of normal power consumption data sliding windows to obtain a trained LSTM autoencoder model; the detection phase comprises: obtaining to-be-detected power consumption data; sliding window sampling the to-be-detected power consumption data to obtain a sequence of detection power consumption data sliding windows with the window length and the sliding step; inputting the sequence of detection power consumption data sliding windows into the trained LSTM autoencoder model to obtain a reconstruction error sequence; based on the reconstruction error sequence, marking time points of abnormal occurrence; wherein training the LSTM autoencoder model based on the sequence of normal power consumption data sliding windows to obtain the trained LSTM autoencoder model comprises: inputting the sequence of normal power consumption data sliding windows into an encoder of the LSTM autoencoder model to obtain a sequence of normal power consumption data time-series context implicit encoding vectors; inputting the sequence of normal power consumption data time-series context implicit encoding vectors into a decoder of the LSTM autoencoder model to obtain a sequence of reconstructed power consumption data sliding windows, comprising: defining a time-series redundancy precursor probability based on the window length and the sliding step; performing local sparsification of the sequence of normal power consumption data time-series context implicit encoding vectors based on the time-series redundancy precursor probability to obtain a local sparsification regulation factor; applying the time-series redundancy precursor probability and the local sparsification regulation factor to feature values of each normal power consumption data time-series context implicit encoding vector in the sequence of normal power consumption data time-series context implicit encoding vectors in a negotiated resolution form to obtain a sequence of normal power consumption data time-series context implicit encoding phase difference coordination vectors; inputting the sequence of normal power consumption data time-series context implicit encoding phase difference coordination vectors into the decoder of the LSTM autoencoder model to obtain the sequence of reconstructed power consumption data sliding windows; calculating a loss function value based on the sequence of normal power consumption data sliding windows and the sequence of reconstructed power consumption data sliding windows; training the LSTM autoencoder model based on the loss function value and through back propagation of gradient descent.

2. The method of claim 1, wherein the step of detecting the local abnormal points comprises the steps of: The encoder of the LSTM autoencoder model is one or more LSTM layers, and the decoder of the LSTM autoencoder model is one or more LSTM layers.

3. The method of claim 1, wherein the step of detecting the local abnormal points comprises the steps of: Calculating a loss function value based on the sequence of normal power consumption data sliding windows and the sequence of reconstructed power consumption data sliding windows comprises: calculating a mean square error of the sequence of normal power consumption data sliding windows and the sequence of reconstructed power consumption data sliding windows as the loss function value.

4. The method of claim 1, wherein the step of detecting the local abnormal points comprises the steps of: Based on the reconstruction error sequence, marking time points of abnormal occurrence comprises: comparing each reconstruction error in the reconstruction error sequence with a preset reconstruction error threshold; If the reconstruction error is greater than the preset reconstruction error threshold, mark the corresponding detection electricity data sliding window as containing a local anomaly; Determine a time point of anomaly occurrence based on the corresponding detection electricity data sliding window.

5. The method of claim 4, wherein the step of detecting the local abnormal points comprises: Determine a time point of anomaly occurrence based on the corresponding detection electricity data sliding window, including marking the timestamp of the center data point of the corresponding detection electricity data sliding window as the time point of anomaly occurrence.

6. A system for detecting local abnormal points of a smart meter, the system comprising: Comprise: The training module comprises: A normal historical electricity data acquisition unit configured to acquire historical electricity data labeled as normal; A normal electricity data extraction unit configured to define a window length and a sliding step, and extract a sequence of normal electricity data sliding windows from the historical electricity data labeled as normal; An LSTM autoencoder training unit configured to train an LSTM autoencoder model based on the sequence of normal electricity data sliding windows to obtain a trained LSTM autoencoder model; The detection module comprises: A to-be-detected electricity data acquisition unit configured to acquire to-be-detected electricity data; A detection electricity data sliding sampling unit configured to perform sliding window sampling on the to-be-detected electricity data with the window length and the sliding step to obtain a sequence of detection electricity data sliding windows; A reconstruction error generation unit configured to input the sequence of detection electricity data sliding windows into the trained LSTM autoencoder model to obtain a sequence of reconstruction errors; An anomaly time point marking unit configured to mark a time point of anomaly occurrence based on the sequence of reconstruction errors; Wherein, training the LSTM autoencoder model based on the sequence of normal electricity data sliding windows to obtain the trained LSTM autoencoder model comprises: Inputting the sequence of normal electricity data sliding windows into an encoder of the LSTM autoencoder model to obtain a sequence of normal electricity data time series context implicit encoding vectors; Inputting the sequence of normal electricity data time series context implicit encoding vectors into a decoder of the LSTM autoencoder model to obtain a sequence of reconstructed electricity data sliding windows, including defining a time series redundancy precursor probability based on the window length and the sliding step, performing local sparsification based on the time series redundancy precursor probability on the sequence of normal electricity data time series context implicit encoding vectors to obtain a local sparsification regulation factor, and applying the time series redundancy precursor probability and the local sparsification regulation factor to feature values of each normal electricity data time series context implicit encoding vector in the sequence of normal electricity data time series context implicit encoding vectors in a negotiated resolution form to obtain a sequence of normal electricity data time series context implicit encoding phase difference coordination vectors, and inputting the sequence of normal electricity data time series context implicit encoding phase difference coordination vectors into the decoder of the LSTM autoencoder model to obtain the sequence of reconstructed electricity data sliding windows; Calculating a loss function value based on the sequence of normal electricity data sliding windows and the sequence of reconstructed electricity data sliding windows; Training the LSTM autoencoder model based on the loss function value and through back propagation of gradient descent.

Images