Electric bicycle charging detection method based on non-intrusive load monitoring

Abstract

The invention relates to the technical field of non-intrusive load monitoring, in particular to an electric bicycle charging detection method based on non-intrusive load monitoring, and the method comprises the steps: pre-constructing a Bi-LSTM network model, and inputting the voltage and current data collected in real time into the Bi-LSTM network model; the Bi-LSTM network model comprises: an acquisition unit; the load identification unit is used for outputting an identification result of whether an electric bicycle charging event exists or not. According to the method, firstly, a non-intrusive load switching monitoring method based on a rule judgment principle and a time threshold value is used to realize accurate monitoring of a switching event; after a switching event is monitored, a bidirectional long-short-term memory network is called, past and future context information is considered at the same time, a long-term dependency relationship and a time sequence mode in a sequence are accurately and comprehensively captured, and the load is effectively and accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of non-invasive load monitoring, and particularly relates to an electric bicycle charging detection method based on non-invasive load monitoring. Background Art

[0002] In recent years, in the face of the explosive growth in the number of electric bicycles, the supporting facilities in many old residential communities are imperfect, and a large number of residents charge their electric vehicles in the corridors or privately pull wires to charge indoors. If electric leakage or spontaneous combustion occurs during the charging of electric bicycles, poisonous gas will spread rapidly along the passage. At the same time, since electric vehicles will block the corridor, residents cannot escape in time. For this reason, technicians in this field have developed a variety of invasive monitoring methods to monitor illegal parking and charging: installing hardware devices such as sensors at each monitored electrical device to directly measure the power consumption of each electrical device. However, due to the problems of high installation and maintenance costs and potential infringement of privacy in the invasive monitoring method, the acceptance of users is not high. Based on this, technicians in this field have further proposed a non-invasive load monitoring method, which provides a new idea for solving the problem; Compared with invasive load monitoring, the NILM non-invasive load monitoring technology collects the total load data of the user side incoming line, obtains the user load usage situation through modeling analysis, and uses the non-invasive load monitoring method to monitor the illegal charging of electric bicycles. Only a monitoring device needs to be installed at the meter box or power bus in the prohibited charging area, and through the identification algorithm, real-time, automatic, and comprehensive monitoring of the illegal parking and charging of electric bicycles can be realized, solving the problems of low efficiency and poor effectiveness of manual inspections.

[0003] The existing non-invasive load monitoring methods mainly include the following: factorial hidden Markov model and its variants, which identify the load characteristics of different devices by modeling time series data; deep learning methods: the NILM non-invasive load monitoring method based on deep neural networks has been widely used, which can handle complex non-linear relationships and improve the accuracy of load identification; graph signal alternating optimization method: using graph signal processing technology to study the load information of residential users to improve the accuracy of total load decomposition; convolutional block attention model: enhancing the model's attention to key features through the attention mechanism to improve the accuracy of load decomposition; method based on image features and transfer learning: this method uses image features, such as the V-I graph, and combines transfer learning technology, showing competitiveness compared with traditional methods.

[0004] Although researchers at home and abroad have done a great deal of work in improving the accuracy of model recognition, enhancing the scalability and robustness of algorithms, etc., the non-intrusive load monitoring technology has not been popularized and used in power companies and on the user side, and there are still many problems to be solved. The main problems include: in the process of load recognition, how to ensure the accuracy of non-intrusive load monitoring in the face of massive electricity consumption load data of different users and the scenario of multiple loads running simultaneously; in terms of switching monitoring, how to accurately capture the transient characteristics at the moment of load switching to achieve fast and accurate load switching recognition; in feature extraction, how to construct a comprehensive and discriminative load feature library to meet the generalization performance requirements of continuously connected new loads, and at the same time ensure the stable and reliable extraction of key load features in a complex electricity consumption environment.

[0005] Based on this, those skilled in the art urgently need to provide a brand-new non-intrusive load monitoring method to solve the problems existing in the above-mentioned prior art. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, and thus provide an electric bicycle charging detection method based on non-intrusive load monitoring.

[0007] An electric bicycle charging detection method based on non-intrusive load monitoring includes the following steps: Pre-construct a Bi-LSTM network model, and input the voltage and current data collected in real time at the meter box or power bus into the Bi-LSTM network model; the Bi-LSTM network model includes: A collection unit for receiving the voltage and current data collected in real time at the meter box or power bus; A switching monitoring unit: using a load switching monitoring method and a preset threshold to detect the voltage and current data to detect whether there is an electrical appliance switching event; A feature extraction unit: when it is detected that there is an electrical appliance switching event, extract the relevant features in the voltage and current data for distinguishing an electric bicycle from other electrical appliance loads; A load recognition unit: according to the pre-constructed load feature database and the cooperation of the Bi-LSTM network, perform load recognition on the relevant features input into the Bi-LSTM network, and output the recognition result of whether there is an electric bicycle charging event; And a prevention response unit: the prevention measure responds to the load recognition result output by the Bi-LSTM network and automatically starts or closes the corresponding prevention operation.

[0008] Preferably, before performing switching monitoring, data quality preprocessing is also included on the voltage and current data, which specifically includes: cleaning processing, noise reduction processing, normalization processing, outlier detection processing, and sampling conversion processing.

[0009] Preferably, the load switching monitoring method and a preset threshold are used to detect the voltage and current data to detect whether there is an electrical appliance switching event, which specifically includes the following steps: Calculate the forward difference sequence or backward difference sequence of the active power; Preliminary event detection: Determine whether the forward difference sequence or backward difference sequence is greater than the power threshold; if the judgment result is yes, output the power value at the moment when the rising event occurs; if the judgment result is no, output the power value at the moment when the falling event occurs: 0; Combine the two sequences respectively containing the power values at the moments when the rising event and the falling event occur; Secondary event detection: Determine whether the rising event sequence or the falling event sequence is greater than the sequence threshold; if the judgment result is yes, output the electrical appliance position at the moment greater than the sequence threshold, and calculate the time difference between adjacent similar events; if the judgment result is no, it is judged as small fluctuations during no-load; Filtering of the time difference between similar events: Further determine whether the time difference between adjacent similar events is less than the time difference threshold based on the electrical appliance position at the moment greater than the sequence threshold output by the secondary event detection; if the judgment result is yes, delete the latter event with a time difference less than the time difference threshold; if the judgment result is no, it is judged as an event without repeated detection; Filtering of the time difference between adjacent different types of events: Calculate the time difference between adjacent different types of event pairs, and determine whether the time difference between adjacent different types of event pairs is less than the time difference threshold between adjacent different types of event pairs; if the judgment result is yes, delete the adjacent different types of event pair, and output the electrical appliance switching event detection result used to describe whether there is an electrical appliance switching event; if the judgment result is no, it is judged as no startup impact event.

[0010] Preferably, when it is detected that there is an electrical appliance switching event, relevant features for distinguishing electric bicycles from other electrical appliance loads are extracted from the voltage and current data; Among them, the relevant features include: active power, reactive power, current amplitude, time-domain waveform, harmonic feature, and V-I trajectory feature.

[0011] Preferably, the network structure of the Bi-LSTM neural network model includes: an input layer for adapting to data in different length ranges - a hidden layer containing two one-dimensional convolutional layers and two Bi-LSTM layers - two fully connected layers with 128 and 1 nodes respectively for result output, which are arranged in sequence.

[0012] 6. The method for detecting electric bicycle charging based on non-intrusive load monitoring according to claim 2, wherein the noise reduction processing includes: wavelet noise reduction and combined filtering; Among them, the wavelet noise reduction processing flow is: wavelet decomposition - threshold calculation - soft threshold processing - wavelet reconstruction; The combined filtering processing flow is: median filtering - zero-phase Butterworth filtering.

[0013] Preferably, for the zero-phase Butterworth filtering, specifically: To avoid the influence of phase shift on power calculation, zero-phase filtering is used. The zero-phase filtering is achieved by applying the Butterworth filter twice, forward and backward, and the cut-off frequency adapts to the characteristic frequency of electric bicycle charging.

[0014] Preferably, before performing relevant feature extraction in the feature extraction unit, the input signal is preprocessed, specifically including: Definition of the input signal: voltage signal: , current signal: , sampling frequency , signal length , time series ; Moving average filtering: Window size , for the signal to be smoothed: ; where represents the value of the signal after smoothing at the discrete time point ; b represents the index offset of the data points within the window; Trend elimination uses a Savitzky-Golay filter: Using a 101-point window and a second-order polynomial fitting to calculate the trend component : ;

[0015] where the coefficient is determined by the least squares method; In the formula: represents the value of the signal after warning trend elimination at the continuous time point t; c represents the order index in the polynomial fitting.

[0016] The technical solution of the present invention has the following advantages: The present invention proposes a Bi-LSTM model for non-intrusive load monitoring based on a bidirectional long short-term memory network to achieve real-time monitoring of illegal parking and charging of electric bicycles. In the Bi-LSTM model, a non-intrusive load switching monitoring method based on the principle of rule judgment combined with a time threshold is first used to accurately monitor switching events. After a switching event is detected, the bidirectional long short-term memory network is called, and considering both past and future context information, it accurately and comprehensively captures the long-term dependencies and temporal patterns in the sequence, effectively and accurately identifying the load. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is the structural block diagram of the Bi-LSTM network model of the present invention; Figure 2 It is the flowchart for switching monitoring of the present invention; Figure 3 It is the structural diagram of the Bi-LSTM neural network used in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0020] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0021] In the description of the present invention, it should be noted that unless otherwise clearly specified and defined, the terms "installation", "connection", and "linkage" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0022] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Embodiment 1 This embodiment discloses an electric bicycle charging detection method based on non-intrusive load monitoring, including the following steps: Pre-build a Bi-LSTM network model and input the voltage and current data collected in real time at the meter box or power bus into the Bi-LSTM network model; as Figure 1 The Bi-LSTM network model includes: A collection unit for receiving the voltage and current data collected in real time at the meter box or power bus; A switching monitoring unit: using a load switching monitoring method and a preset threshold to detect the voltage and current data to detect whether there is an electrical appliance switching event; A feature extraction unit: when detecting an electrical appliance switching event, extract the relevant features in the voltage and current data for distinguishing an electric bicycle from other electrical appliance loads; A load identification unit: according to the pre-built load feature database and the Bi-LSTM network, cooperate to identify the relevant features input into the Bi-LSTM network, and output the identification result of whether there is an electric bicycle charging event; And a prevention response unit: the prevention measure responds to the load identification result output by the Bi-LSTM network and automatically starts or closes the corresponding prevention operation.

[0024] In this embodiment, before performing switching monitoring, data quality preprocessing is also included for the voltage and current data, specifically including: cleaning processing, noise reduction processing, normalization processing, outlier detection processing, and sampling conversion processing. Specifically: 1. Data loading and definition Let the original signal be a time series: ; where the sampling frequency ; 2. Data cleaning: 2.1 Linear interpolation to fill in missing values: For the time points of missing values in the time series , interpolation is performed using the nearest valid points before and after and : ; where , continuous filling of missing values is achieved; 2.2 Deletion of abnormal segments Define a sliding window . If the number of consecutive missing points within the window : ; Delete the section marked as NaN; NAN represents the exponential function. When x(t) is NaN, which is not a number, the function value is 1, otherwise it is 0.

[0025] 3. Wavelet denoising 3.1 Wavelet decomposition Select the Daubechies-4 wavelet for 5-level decomposition:

[0026] 3.2 Threshold calculation ; ; 3.3 Soft threshold processing ; 3.4 Wavelet reconstruction .

[0027] 4. Combined filtering 4.1 Median filtering, specifically using anti-impulse noise: ; 4.2 Zero-phase Butterworth filtering To avoid the influence of phase shift on power calculation, zero-phase filtering is used. Zero-phase filtering is achieved by applying the Butterworth filter twice, forward and backward: ; Implemented by the difference equation: ; Coefficient is generated by the function of the Butterworth filter , and the cut-off frequency Adapt to the characteristic frequency of electric bicycle charging; r represents the discrete-time index, i.e., the sequence number of the signal sampling point currently being processed; forward and backward represent the forward filtering direction and the backward filtering direction respectively; 5. Anomaly Detection 5.1 Feature Engineering ; Feature matrix ; 5.2 Anomaly Scoring Use Mahalanobis distance: ; ; 6. Normalization Processing Improved to Z-score standardization to retain distribution information: ; In the formula: Represents the signal value after normalization processing; Represents the original signal Mean value; Represents the original signal Standard deviation; 7. Resampling 7.1 Anti-aliasing Filtering First perform low-pass filtering to avoid aliasing: ; 7.2 Polynomial Interpolation ; The sinc function realizes ideal interpolation.

[0028] Explanation of Formula Symbols

[0029] Such as Figure 2 Use the load switching monitoring method and the preset threshold to detect the voltage and current data to detect whether there is an electrical switching event, which specifically includes the following steps: Calculate the forward difference sequence or backward difference sequence of the active power: ; ; Preliminary event detection: Determine whether the forward difference sequence or the backward difference sequence is greater than the power threshold; if the judgment result is yes, output the power value at the moment when the rising event occurs; if the judgment result is no, output the power value at the moment when the falling event occurs: 0; Merge two sequences respectively containing the power values at the moments when the rising event and the falling event occur. Define the event marking function: ; ; Merge the event sequences: ; Secondary event detection: Determine whether the rising event sequence or the falling event sequence is greater than the sequence threshold; if the judgment result is yes, output the electrical appliance position at the moment greater than the sequence threshold, and calculate the time difference between adjacent similar events; if the judgment result is no, judge it as small fluctuations during no-load. ; Filter the time difference of similar events: On the basis of the electrical appliance position output by the secondary event detection at the moment greater than the sequence threshold, further determine whether the time difference between adjacent similar events is less than the time difference threshold; if the judgment result is yes, delete the latter event with a time difference less than the time difference threshold; if the judgment result is no, judge it as an event without duplicate detection; Filter the time difference of similar events (rising or falling).

[0030] Let the time series of similar events be , calculate the adjacent time difference: ; ; Filtering condition: ; If , delete the latter event: Delete ).

[0031] Filter the time difference of adjacent different types of events: Calculate the time difference of adjacent different types of event pairs, and determine whether the time difference of adjacent different types of event pairs is less than the time difference threshold of adjacent different types of event pairs; if the judgment result is yes, delete the adjacent different types of event pair, and output the electrical appliance switching event detection result used to describe whether there is an electrical appliance switching event; if the judgment result is no, judge it as no start-up impact event.

[0032] Define the time difference of adjacent rising-falling event pairs as: ; Filtering condition: ; If , delete this event pair.

[0033] The final output of the remaining event set is: ;

[0034] Furthermore, when detecting electrical switching events, extract the relevant features in the voltage and current data to distinguish electric bicycles from other electrical loads; Among them, the relevant features include: active power, reactive power, current amplitude, time-domain waveform, harmonic features, and V-I trajectory features. Specifically: Before performing relevant feature extraction in the feature extraction unit, first preprocess the input signal, specifically including: 1. Input signal definition: Voltage signal: , current signal: , sampling rate , signal length H, time series ; Moving average filtering: Window size , smooth the signal : ; In the formula, represents the value of the signal after smoothing at the discrete time point ; b represents the index offset of the data points within the window; Trend elimination uses a Savitzky-Golay filter: Adopt a 101-point window and a second-order polynomial fit to calculate the trend component :

[0035]

[0036] .

[0037] 2. Basic power features Apparent power : ; Active power : ; Reactive power : ; Power factor : ; 3. Current characteristics RMS value: ; Peak current:

[0038] Peak count: Define the peak as a local maximum that satisfies the following conditions: ; Total number of peaks: ; Current skewness: ; 4. Time-domain waveform characteristics Sliding window standard deviation: Window size corresponding to 1 cycle of a 50 Hz signal: ; Statistic: ; Current kurtosis

[0039] 5. Harmonic characteristics: Fast Fourier transform: Current signal spectrum: ; Magnitude spectrum:

[0040] Fundamental wave extraction: Fundamental wave frequency and magnitude : ; Total harmonic distortion: Extract the magnitudes of the 2nd - 20th harmonics :

[0041] Ratio of odd harmonics:

[0042] 6. V-I trajectory characteristics Phase difference calculation: Instantaneous phase angle: ; Average phase difference: ; Ellipse fitting covariance matrix: ; Eigenvalue : ; Symbol description

[0043] According to the pre - constructed load characteristic database and the Bi - LSTM network, the relevant characteristics input into the Bi - LSTM network are used for load identification, and the identification result of whether there is an electric bicycle charging event is output, specifically including: 1. Input definition After feature extraction, the feature vector at each time step is: ; : Feature dimension; : Time series length.

[0044] 2. Bi - LSTM network structure Bi - LSTM is composed of a forward LSTM and a backward LSTM. Each LSTM cell contains the following calculation modules: 2.1. Forward LSTM Input gate, controlling the input of new information: ; Forget gate, controlling the forgetting of old information: ; Candidate cell state, generating new candidate values: ; Cell state update, combining forgetting and input: ; Output gate, controlling the amount of information output: ; Hidden state output: ; 2.2. Backward LSTM The backward LSTM starts from to The reverse processing sequence, the formula is symmetric to the forward LSTM: ; 2.3 Bidirectional output concatenation The final hidden state at each time step is formed by concatenating the forward and backward hidden states: ; Where: : The number of hidden units of a single LSTM; : The output of the Bi-LSTM, used for subsequent classification.

[0045] 3. Classification layer Map the Bi-LSTM output to the classification result: ; Where: : The weight matrix of the fully connected layer; : The bias vector; : The number of classes, such as electric bicycle charging / other loads.

[0046] Symbol description

[0047] In this embodiment, the network structure of the Bi-LSTM neural network model includes, in sequence: an input layer for adapting to data of different length ranges - a hidden layer containing two one-dimensional convolutional layers and two Bi-LSTM layers - two fully connected layers with 128 and 1 nodes respectively for result output.

[0048] Specifically: For example Figure 3 In this embodiment, the input of the input layer is one-dimensional sequence data, which can better adapt to data of different length ranges. The hidden layer is set to two one-dimensional convolutional layers and two Bi-LSTM layers. Each one-dimensional convolutional layer uses a stride of 1 and 16 convolutional kernels of size 4. A linear activation layer is set after the one-dimensional convolutional layer for feature extraction of the data. Then it passes through two Bi-LSTM layers with 64 and 128 nodes respectively, which are the core of the model. Finally, two fully connected layers with 128 and 1 nodes respectively are used for result output. A Tanh activation function is added after the fully connected layer with 128 nodes to synthesize all features and input them into the next fully connected layer.

[0049] In addition, the software used in this embodiment is Pycharm software, and the deep learning framework is the PyTorch framework. The dataset used in this embodiment is the UK-DALE dataset. UK-DALE, full name UK Domestic Appliance-Level Electricity dataset, is an open-source household electricity consumption monitoring benchmark dataset launched by the Jack Kelly team of Imperial College London in 2014, collecting electricity consumption data from five households. Since the data of the first household contains 53 types of household appliance load types and has a time span of up to 655 days, which is the household with the longest collection time in the dataset and is more representative and complex. Therefore, the data of the first household is selected as the training set of the model in this embodiment. The high-frequency waveform layer in the dataset uses a sampling rate of 16KHz, which can accurately capture the high-frequency transient characteristics at the moment when the electrical appliance starts and stops. The low-frequency data uses a sampling rate of 1 / 6Hz, and the active power time series data of each independent electrical appliance is recorded at intervals of 6 seconds. At the same time, for the total power and apparent power, a sampling rate of 1Hz is used, which can construct the dynamic curve of the total household load.

[0050] Table

[0051] The test results show that: the indicators of the accuracy, recall rate and F1 score of this model are all higher than those of LSTM and CNN, which indicates that the performance of the model provided by the present invention is superior to that of traditional algorithms. Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for detecting electric bicycle charging based on non-intrusive load monitoring, characterized in that, It includes the following steps: Pre-build a Bi-LSTM network model and input the voltage and current data collected in real time at the meter box or power bus into the Bi-LSTM network model; Bi -LSTM network model includes: An acquisition unit for receiving the voltage and current data collected in real time at the meter box or power bus; A switching monitoring unit: detecting the voltage and current data by using a load switching monitoring method and a preset threshold to detect whether there is an electrical appliance switching event; A feature extraction unit: when detecting an electrical appliance switching event, extracting relevant features in the voltage and current data for distinguishing an electric bicycle from other electrical appliance loads; A load identification unit: performing load identification on the relevant features input into the Bi-LSTM network according to the pre-built load feature database and the Bi-LSTM network, and outputting an identification result of whether there is an electric bicycle charging event; And a prevention response unit: the prevention measure responds to the load identification result output by the Bi-LSTM network and automatically starts or closes the corresponding prevention operation.

2. The electric bicycle charging detection method based on non-intrusive load monitoring according to claim 1, wherein, Before performing switching monitoring, data quality preprocessing is also included for the voltage and current data, specifically including: cleaning processing, noise reduction processing, normalization processing, outlier detection processing, and sampling conversion processing.

3. The electric bicycle charging detection method based on non-intrusive load monitoring according to claim 1, wherein Detecting the voltage and current data by using a load switching monitoring method and a preset threshold to detect whether there is an electrical appliance switching event, specifically including the following steps: Calculating the forward difference sequence or backward difference sequence of the active power; Preliminary event detection: judging whether the forward difference sequence or backward difference sequence is greater than the power threshold; if the judgment result is yes, outputting the power value at the moment when the rising event occurs; if the judgment result is no, outputting the power value at the moment when the falling event occurs: 0; merging two sequences respectively containing the power values at the moments when the rising event and the falling event occur; Secondary event detection: judging whether the rising event sequence or the falling event sequence is greater than the sequence threshold; if the judgment result is yes, outputting the electrical appliance position at the moment greater than the sequence threshold and calculating the time difference between adjacent similar events; if the judgment result is no, judging it as small fluctuations during no-load; Filtering of the time difference between adjacent similar events: further judging whether the time difference between adjacent similar events is less than the time difference threshold on the basis of the electrical appliance position at the moment greater than the sequence threshold output by the secondary event detection; if the judgment result is yes, deleting the latter event with a time difference less than the time difference threshold; if the judgment result is no, judging it as an event without repeated detection; Filtering of the time difference between adjacent different types of events: calculating the time difference between adjacent different types of event pairs and judging whether the time difference between adjacent different types of event pairs is less than the time difference threshold between adjacent different types of event pairs; if the judgment result is yes, deleting the adjacent different types of event pair and outputting the electrical appliance switching event detection result for describing whether there is an electrical appliance switching event; if the judgment result is no, judging it as no startup impact event.

4. The electric bicycle charging detection method based on non-intrusive load monitoring according to claim 1, wherein, When detecting an electrical appliance switching event, extracting relevant features in the voltage and current data for distinguishing an electric bicycle from other electrical appliance loads; Among them, the relevant features include: active power, reactive power, current amplitude, time-domain waveform, harmonic feature, and V-I trajectory feature.

5. A method for detecting electric bicycle charging based on non-intrusive load monitoring according to claim 1, characterized in that The network structure of the Bi-LSTM neural network model includes, in sequence: an input layer for adapting to data in different length ranges, a hidden layer containing two one-dimensional convolutional layers and two Bi-LSTM layers, and two fully connected layers with 128 and 1 nodes respectively for result output.

6. The electric bicycle charging detection method based on non-intrusive load monitoring according to claim 2, wherein The noise reduction processing includes: wavelet noise reduction and combined filtering; Among them, the wavelet noise reduction processing flow is: wavelet decomposition - threshold calculation - soft threshold processing - wavelet reconstruction; The combined filtering processing flow is: median filtering - zero-phase Butterworth filtering.

7. The electric bicycle charging detection method based on non-intrusive load monitoring according to claim 6, wherein For zero-phase Butterworth filtering, specifically: To avoid the influence of phase shift on power calculation, zero-phase filtering is used. Zero-phase filtering is achieved by applying the Butterworth filter twice, forward and backward, and the cut-off frequency adapts to the characteristic frequency of electric bicycle charging.

8. A method for detecting electric bicycle charging based on non-intrusive load monitoring according to claim 4, characterized in that, Before performing relevant feature extraction in the feature extraction unit, the input signal is preprocessed, specifically including: Input signal definition: Voltage signal: , Current signal: , Sampling frequency , Signal length , Time series ; Moving average filtering: Window size , smooth the signal as follows: ; wherein, represents the value of the signal after smoothing at the discrete time point ; b represents the index offset of the data points within the window; Trend elimination uses a Savitzky-Golay filter: Using a 101-point window and a second-order polynomial fitting to calculate the trend component : ； Among them, the coefficient is determined by the least squares method; In the formula: represents the value of the signal after warning trend elimination at consecutive time points t; c represents the order index in polynomial fitting.

Images