Non-intrusive load monitoring system based on deep residual network

Abstract

The application discloses a non-invasive load monitoring system based on a deep residual network, and relates to the technical field of power monitoring.The non-invasive load monitoring system comprises a self-recovery module, a mapping module and a verification module.The self-recovery module blocks the signal transmission of the disturbed frequency band in the preset identification structure according to the power grid quality evaluation report, compensates the response gain of the pure frequency band, and outputs an environment self-recovery type parameter.The mapping module, on the basis of loading the environment self-recovery type parameter, responds to the load action event triggered by the total current signal, intercepts the corresponding power fluctuation data, maps the power fluctuation data to a polar coordinate system, and outputs a space-time trajectory load state matrix.The verification module performs real-time logical verification on the response state of the space-time trajectory load state matrix under the electromagnetic physical constraint, and outputs a load identification probability matrix.The application improves the anti-interference ability of the detection system under complex and changeable power grid interference, and effectively guarantees the high credibility of the electric appliance type and energy consumption diagnosis report in the physical reality level.

Description

Technical Field

[0001] This invention relates to the field of power monitoring technology, and in particular to a non-intrusive load monitoring system based on deep residual networks. Background Technology

[0002] Non-intrusive load monitoring (NILM), a core technology for demand-side management in smart grids, has seen significant development in recent years thanks to the evolution of deep learning algorithms. Existing technologies often employ high-performance models such as deep residual networks (ResNet) to learn high-frequency or transient energy indicators at the bus terminal, establishing a mapping relationship between steady-state load characteristics and specific appliance types. This allows for refined identification of terminal loads without intruding on the user's internal network. These methods greatly improve the feature extraction depth and classification convergence speed of multi-objective composite loads in complex scenarios, becoming a current research hotspot in the field of energy consumption behavior sensing.

[0003] However, in actual industrial or residential power distribution networks, the widespread presence of nonlinear loads often leads to dynamic harmonic distortion and background noise interference. For example, the presence of high-power variable frequency air conditioners, LED lighting clusters, or rectifiers in the power grid injects a large amount of high-order harmonics and unpredictable random burst noise into the bus. Existing deep residual network monitoring systems rely excessively on high-dimensional data-driven approaches and lack adaptive adjustment mechanisms for complex power grid environments and constraints from underlying electrical and physical laws. This results in the network's deep convolutional kernels being unable to distinguish between transient features generated by electrical actions and pseudo-features superimposed on background noise when facing the aforementioned strong interference frequency bands. This easily leads to confusion in neuronal activation states and response distortion, severely impacting the accuracy and robustness of load identification. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a non-intrusive load monitoring system based on deep residual networks to solve the problem of low load identification accuracy caused by model confusion and spurious features under strong interference due to the lack of environmental adaptation and physical law constraints.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a non-intrusive load monitoring system based on deep residual networks, comprising: The acquisition module acquires the bus voltage signal and the total current signal, performs phase anchoring on the total current signal based on the voltage zero crossing point, obtains the basic power data set, extracts background power quality data, and generates a power grid quality assessment report. The self-healing module blocks the signal transmission of the interfered frequency band in the preset identification structure based on the power grid quality assessment report, compensates for the response gain of the clean frequency band, and outputs environmental self-healing parameters. The mapping module, based on the self-healing parameters of the loaded environment, responds to the load action event triggered by the total current signal, extracts the corresponding power consumption fluctuation data, maps it to the polar coordinate system, and outputs the spatiotemporal trajectory load state matrix. The verification module performs real-time logical verification of the response state of the spatiotemporal trajectory load state matrix under electromagnetic physical constraints and outputs a load identification probability matrix. The diagnostic module identifies the type of electrical appliance based on the load identification probability matrix, and outputs the load monitoring results when the measured power of the total current signal and the total power of the identified sub-loads satisfy the energy conservation constraint.

[0007] Preferably, the method for obtaining the basic electricity dataset includes: Based on the preset hardware magnetic saturation correction parameters, the total current signal is subjected to time-domain phase compensation, and the phase-compensated total current signal and the acquired bus voltage signal are synchronously buffered to extract the original sampling time sequence with a fixed period. Lock the zero-crossing point and peak point of the bus voltage signal, and extract the relative time difference between each sampling point in the total current signal and the zero-crossing point of the waveform to generate phase space labels; Phase space labels are mapped and bound to the total current signal, and the bound total current signal, bus voltage signal, and waveform zero-crossing points and waveform peak points are combined to obtain the basic power data set.

[0008] Preferably, the method for generating the power grid quality assessment report includes: The bus voltage signal is subjected to frequency domain mapping transformation to obtain the corresponding voltage spectrum feature distribution. The total harmonic distortion rate, the proportion of odd harmonic energy and the statistical characteristics of voltage fluctuation are extracted from the voltage spectrum feature distribution and combined into background power quality data. The background power quality data is compared with the preset standard to identify the interference frequency band that exceeds the preset standard. The interference frequency band is defined as the feature suppression area. The background power quality data is entered into the real-time constructed environmental background noise fingerprint database. The signal-to-noise ratio prediction value and frequency confidence in the environmental background noise fingerprint database are extracted. The feature suppression region, signal-to-noise ratio prediction value, and frequency confidence are jointly and structurally encapsulated to generate a multi-dimensional power grid environmental quality assessment report.

[0009] Preferably, the method for blocking signal transmission in the interfered frequency band of the preset identification structure includes: The purity level of the current power grid environment is determined by the predicted signal-to-noise ratio and frequency confidence level. Based on the purity level, a topology control command is triggered to initialize a control feature vector corresponding to the number of feature channels of the preset recognition structure. Set the feature bit values ​​of the frequency band features corresponding to the feature suppression region in the control feature vector to zero, generate a topology control mask, and use the topology control mask to map the feature extraction channel corresponding to the feature suppression region in the preset recognition structure. By using a topology control mask to truncate the propagation weights of the feature extraction channel to zero, the extension and propagation of features in the interfered frequency band within the preset recognition structure is blocked.

[0010] Preferably, the method for compensating the response gain of the clean frequency band includes: Extract the frequency range not covered by the feature suppression region in the multi-dimensional power grid environmental quality assessment report, obtain the high signal-to-noise ratio frequency band, and locate the feature path corresponding to the high signal-to-noise ratio frequency band in the preset identification structure as the pure channel to be compensated. The gain control quantity is obtained based on the redundant computing resources released after blocking the interfered frequency band, and the scaling parameters inside the clean channel to be compensated are dynamically adjusted using the gain control quantity. The zero-truncated propagation weights and scaling parameters are recombined as a whole, the global weight matrix of the preset recognition structure is updated, and the recombined global weight matrix is ​​solidified and output as environmental self-healing parameters.

[0011] Preferably, the method for responding to a load action event triggered by a total current signal includes: The environmental self-healing parameters are loaded into the preset identification structure, so that the preset identification structure operates in real-time monitoring state under the configuration of the environmental self-healing parameters. Under real-time monitoring, the effective value change rate of the total current signal and the real-time power increment are extracted in real time, and a dynamic action threshold is established based on the characteristic suppression area in the multi-dimensional power grid environmental quality assessment report. The effective rate of change of current or real-time power increment is compared with the dynamic action threshold. When the dynamic action threshold is exceeded, it is determined that the total current signal has triggered a load action event.

[0012] Preferably, the method for outputting the spatiotemporal trajectory load state matrix includes: In response to load action events, backtrack and extract power consumption fluctuation data with a fixed number of cycles, and extract transient voltage sequences with the same sampling period as the power consumption fluctuation data based on the zero crossing point and peak point of the waveform. By mapping the transient voltage sequence as the horizontal axis parameter and the power consumption fluctuation data as the vertical axis parameter to the polar coordinate system, a voltage-current trajectory curve is generated. The voltage-current trajectory curve is denoised and enhanced by using the feature suppression region in the multidimensional power grid environmental quality assessment report. The enhanced voltage-current trajectory curve is then extracted by discretized grid spatial mapping and pixel mapping to output the spatiotemporal trajectory load state matrix.

[0013] Preferably, the method for outputting the load identification probability matrix includes: The spatiotemporal trajectory load state matrix is ​​input into a preset recognition structure for multi-scale feature extraction, resulting in multi-scale activation feature maps output by multiple internal hidden layers. Based on Kirchhoff's current law and the energy continuity equation, a real-time electromagnetic physical constraint review is performed on the multi-scale activation feature map to determine whether there are abnormal jumps in the electrical activation components of the multi-scale activation feature map. When an abnormal jump is detected, the contribution weight of the feature propagation path is reduced to a state where it cannot contribute to the forward propagation of subsequent feature extraction. The original phase information in the basic power dataset is reorganized into the multi-scale activation feature map using the jump connection structure of the preset recognition structure for phase alignment correction, and the load identification probability matrix is ​​output.

[0014] Preferably, the method for identifying the type of electrical appliance based on the load identification probability matrix includes: Extract the identification probability value corresponding to each electrical appliance type in the load identification probability matrix, and output the candidate load component set when the identification probability value is greater than a fixed probability threshold; Obtain the standard electrical power fingerprint and standard phase features corresponding to the candidate load component set in the electrical appliance feature library; match and verify the standard electrical power fingerprint and standard phase features with the basic power dataset to identify the electrical appliance type.

[0015] Preferably, the method for outputting load monitoring results includes: Obtain the standard electrical power fingerprint to synthesize the expected simulated total current waveform, and extract the total power of the sub-loads from the expected simulated total current waveform; The measured power of the total current signal extracted from the basic electricity dataset is compared with the total power of the sub-loads to obtain the energy deviation residual. When the energy deviation residual is within the allowable range of the electricity metering standard, it is determined that the energy conservation constraint is met. When the residual energy deviation exceeds the allowable range of the electricity metering standard, the bias parameters of the output layer of the preset identification structure are adjusted by the feedback command, and the difference waveform obtained by subtracting the expected simulated total current waveform from the total current signal point by point is fed back to the preset identification structure for secondary feature extraction to correct the identification of the electrical appliance type. While satisfying the energy conservation constraint, the system integrates and identifies the equipment name, operating mode, energy consumption, and electricity health recommendations corresponding to the type of electrical appliance, and outputs load monitoring results.

[0016] The beneficial effects of this invention are as follows: by extracting background power quality data and generating a power grid environmental quality assessment report, the noise distribution characteristics and purity level of the current complex power grid environment on the user side are accurately quantified and determined. This is applied to the monitoring front end to lock harmonic interference in order to provide topology control, enabling the deep learning network to have hardware-level environmental adaptation and noise resistance self-healing capabilities, and improving the robustness of operation in harsh power grid environments. By performing real-time physical constraint review on the activated feature map and triggering feedback adjustment and secondary feature extraction based on power deviation residuals at the output end, the probabilistic prediction of deep learning and the deterministic classical electromagnetic physics laws are deeply integrated and closed-loop verified at the feature layer and output layer. This is applied to the identification and decision-making stage to conduct compliance review and reverse self-correction of the output results to eliminate false identifications, ensuring that the diagnostic report meets the energy conservation constraint and reducing the false alarm rate. The combination of these two approaches enhances the anti-interference capability of the detection system under complex and variable power grid interference, effectively ensuring the high credibility of the appliance type and energy consumption diagnostic report at the physical reality level, and making the monitoring results closer to the actual energy conservation state. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the non-intrusive load monitoring system based on deep residual networks in this invention.

[0019] Figure 2 This is a flowchart of the output environment self-healing parameters in this invention.

[0020] Figure 3 This is a flowchart of the output load monitoring results in this invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figure 1 , Figure 2 and Figure 3 As one embodiment of the present invention, this embodiment provides a non-intrusive load monitoring system based on deep residual networks, comprising the following steps: Methods for obtaining basic electricity datasets include: The phase lead or lag compensation angle corresponding to different current amplitudes stored in non-volatile memory is used as the preset hardware magnetic saturation correction parameter. For example, the compensation angle is set to two degrees. When the total current signal collected by the current transformer is distorted in the time domain and phase lags due to the core iron core entering the magnetic saturation state, the sampling amplitude of the current total current signal is read and the matching phase compensation angle is retrieved from the preset hardware magnetic saturation correction parameter. The time axis of the currently collected total current signal is shifted to achieve time domain phase compensation. Under the trigger of the same high-frequency sampling clock source, the phase-compensated total current signal and the bus voltage signal collected by the voltage transformer are timestamped from the same source and sent to the static random access memory to achieve synchronous buffering. In the static random access memory, a cyclic sliding window with a window length corresponding to the fixed cycle of the power grid frequency is set (for example, the power frequency example value is 50Hz and the fixed cycle example value is 20 complete cycles, i.e., the example value is 0.4 seconds). When the amount of data in the synchronous buffer reaches the number of sampling points corresponding to 0.4 seconds, the phase-compensated total current signal and the collected bus voltage signal in the cyclic sliding window are intercepted and read, thereby completing the interception of the original sampling time sequence of the fixed cycle.

[0025] Based on the bus voltage signal contained in the original sampling time series, the first-order and second-order difference operations of subtracting adjacent points are performed on the discrete sampling points of the continuous bus voltage signal to calculate the voltage slope and rate of change of each sampling interval, thereby extracting the differential rate of change. When the value of a bus voltage signal sampling point changes from negative to positive or from positive to negative and the first-order difference value is not zero, the time corresponding to the sampling point is determined to be the zero-crossing point of the voltage waveform. When the first-order difference value changes from positive to negative or from negative to positive and the sign of the second-order difference value changes, the corresponding extreme sampling point is determined to be the peak point of the waveform; thus locking the zero-crossing point and peak point of the bus voltage signal. The total current signal after time-domain phase compensation is retrieved within the same time window. Using the time reference of the high-frequency sampling clock source, the difference between the timestamp of each current sampling point in the total current signal and the timestamp of the zero-crossing point of the previous adjacent locked waveform is calculated. Each discrete time difference is used as the relative time difference of each corresponding current sampling point. The relative time difference is encapsulated into a one-dimensional feature vector to generate a phase space label.

[0026] Based on the timestamp index of the high-frequency sampling clock source, each one-dimensional feature vector in the phase space label is written into the data structure of each current sampling point in the total current signal after time-domain phase compensation in the form of an append field, so that each current sampling point contains its own relative time difference information, thus completing the mapping and binding of the phase space label to the total current signal. Within the continuous data storage space of this structure, the total current signal mapped and bound with phase space labels, the bus voltage signal from the original sampling time series with a fixed period extracted in the previous stage, and the timestamp data of the waveform zero crossing point and the timestamp data of the waveform peak point are matrix spliced ​​and multi-channel composite encapsulated according to the chronological order of occurrence to obtain the basic power dataset.

[0027] Methods for generating power grid quality assessment reports include: Based on the bus voltage signal in the basic power data set, the time-domain form of the bus voltage signal is mapped and transformed to the frequency domain to obtain a discrete complex spectrum; the square of the modulus of the complex at each frequency point is calculated to obtain the power spectral density, and the corresponding voltage spectrum characteristic distribution is obtained. In the voltage spectrum characteristic distribution, the fundamental frequency point is located according to the power grid reference frequency, and the energy accumulation value of the fundamental frequency point and its neighborhood is extracted from the corresponding power spectral density to obtain the total fundamental energy value. At the same time, each harmonic frequency point is located according to an integer multiple of the power grid reference frequency, and the power spectral density value corresponding to each harmonic frequency point is integrated and accumulated to obtain the total harmonic energy value. The total fundamental energy value and the total harmonic energy value are retrieved, and the ratio of the sum of the root mean square values ​​of each harmonic component to the root mean square value of the fundamental component is calculated to obtain the total harmonic distortion rate. At the same time, the energy values ​​of the 3rd, 5th, 7th and 9th harmonics within a specific frequency range (example values ​​are 1kHz to 5kHz) are screened out separately, and their percentage of the total energy is calculated to obtain the energy proportion of odd harmonics. The variance and standard deviation of the root mean square value of the bus voltage signal for multiple consecutive cycles are statistically analyzed to obtain the statistical characteristics of voltage fluctuation that characterize the degree of flicker. The total harmonic distortion rate, the proportion of odd harmonic energy, and the statistical characteristics of voltage fluctuations are cascaded into one-dimensional vectors and combined to form background power quality data.

[0028] The total harmonic distortion rate and the proportion of odd harmonic energy in the background power quality data are compared with the voltage harmonic limit (the limit is 5% in the example) in the pre-configured power quality public grid harmonic industry standard. The values ​​are subtracted one by one. When the proportion of odd harmonic energy in a specific frequency range (the example value is 1kHz to 5kHz) is greater than 5%, the frequency range is determined to be an interference band that exceeds the preset standard. The Boolean value of the frequency point corresponding to the interference band is logically set to 1 in the spectrum mask vector, and the interference band is defined as a characteristic suppression region. The current background power quality data is appended to the real-time constructed environmental background noise fingerprint database in the form of key-value pairs with the current timestamp for rolling storage. The background power quality data of historical time periods in the environmental background noise fingerprint database is used to perform time series averaging calculation. The ratio of the fundamental signal energy to the total residual noise energy after removing the feature suppression region is calculated to obtain the signal-to-noise ratio prediction value. Simultaneously, all background power quality data stored in the environmental background noise fingerprint database within a preset historical observation period (the preset historical observation period is an example of 24 hours) are retrieved, and the total number of times the proportion of odd harmonic energy in a specific interference frequency band exceeds the preset standard (the preset standard is an example of 5%) is counted. The total number of times the standard exceeds the standard is divided by the total number of observation periods within 24 hours to obtain the recurrence probability distribution of the repeated triggering of the standard state in the historical time domain of the specific interference frequency band. This probability distribution is used as the frequency confidence level. The environmental background noise fingerprint database is constructed in real time by continuously receiving the background power quality data calculated and output in each cycle, using the absolute timestamp provided by the current high-frequency sampling clock source as the key, and using the background power quality data containing total harmonic distortion rate, odd harmonic energy ratio and voltage fluctuation statistical characteristics as the value, and dynamically adding and storing the data in real time according to the first-in-first-out coverage rule.

[0029] Based on the preset field mapping rules, a one-to-one mapping relationship is established between the power characteristic parameters of different dimensions and the exclusive text labels in the text data; For example: mapping the frequency band boundary values ​​contained in the feature suppression region to the "interference frequency band range" label in the text data; Map the predicted signal-to-noise ratio (SNR) values ​​to the “system SNR” label in the text data; Map frequency confidence scores to “probability of occurrence” labels in the text data; Subsequently, a digital signal processor is used to treat the frequency band boundary values, signal-to-noise ratio prediction values, and frequency confidence scores defined as feature suppression regions in the spectral mask vector as independent labeled data items, and write them sequentially into the corresponding text label slots according to the preset field mapping rules to generate a multi-dimensional power grid environmental quality assessment report.

[0030] Traditional non-intrusive load monitoring systems often suffer from severe distortion of the acquired current signal in specific frequency bands due to hardware magnetic saturation or sudden harmonic interference when facing complex power grid environments. Deep models, lacking environmental awareness, amplify noise signals through layer-by-layer convolution, significantly reducing the accuracy of load identification. Therefore, this invention reconstructs the underlying topology of the identification network in real time by evaluating power grid power quality, thereby achieving interference cutoff and adaptive gain compensation for clean signals at the algorithm level. The specific steps are as follows: Before the forward propagation operation begins, the predefined recognition structure is first defined as a deep residual network. This structure consists of an input layer, a first convolutional layer, multiple interconnected residual blocks with skip connections, and a final fully connected layer. To meet the requirements of multi-scale parallel feature extraction of the spatiotemporal trajectory load state matrix, the backbone topology of the deep residual network is cascaded into three layers: a front shallow residual block (specifically, from the first convolutional layer to the fourth residual block in this example), a middle residual block (specifically, from the fifth residual block to the eighth residual block in this example), and a rear deep residual block (specifically, from the ninth residual block to the fully connected layer in this example). Each first convolutional layer or specific residual block contains a predetermined number of feature extraction channels for spatial convolution of the power characteristics mapped to different frequency bands; the example number of feature extraction channels is 64.

[0031] Methods for blocking signal transmission in the interfered frequency band of the preset identification structure include: The predicted signal-to-noise ratio (SNR) value is compared with the preset SNR grading threshold (the example of the preset SNR grading threshold is 30dB), and the frequency confidence is compared with the preset interference recurrence rate threshold (the example of the preset interference recurrence rate threshold is 15%). When the predicted signal-to-noise ratio is greater than 30dB and the frequency confidence is less than 15%, the purity level of the current power grid environment is determined to be high purity level, with an example value of level 1. When the predicted signal-to-noise ratio (SNR) is less than or equal to 30 dB and the frequency confidence level is greater than or equal to 15%, the purity level of the current power grid environment is determined to be low purity level. For example, the value is level 3. The purity level of the current power grid environment is determined by the predicted SNR and the frequency confidence level.

[0032] The internal control mapping table is retrieved based on the purity level. When the purity level is low (level 3), a conditional branch instruction is executed to trigger the topology control instruction. Obtain the number of feature channels in the first convolutional layer or a specific residual block in the deep residual network. For example, the number of feature channels is 64. Allocate a set of contiguous storage units of length 64 in the running memory and assign an initial value of 1 to each feature bit in the contiguous storage unit, thereby initializing a control feature vector corresponding to the number of feature channels of the preset recognition structure.

[0033] Identify the frequency band range corresponding to the feature suppression region (example value is within 1kHz to 5kHz), find the specific feature bits in the control feature vector that are responsible for processing the frequency band features within 1kHz to 5kHz, such as the 5th to 9th feature bits, and directly modify the values ​​of the 5th to 9th feature bits in the control feature vector from the initial value of 1 to 0 through memory rewrite instructions, while keeping the values ​​of the remaining feature bits as 1, thereby generating the topology control mask; Establish a one-to-one correspondence pointer link between each feature bit in the topology control mask and the feature extraction channel of each convolution kernel in the deep residual network, so that the feature bits with a value of 0 in the topology control mask accurately point to and map the feature extraction channel corresponding to the feature suppression region in the preset recognition structure.

[0034] During the forward propagation feature extraction process of the preset recognition structure, namely the deep residual network, when the input basic electricity dataset flows through the feature extraction channel corresponding to the feature suppression region, the dot product matrix operation operator is called to perform channel-by-channel multiplication of the convolution kernel propagation weight matrix of each feature extraction channel with the value of the corresponding feature bit in the topology control mask. Since the feature bit value of the frequency band feature corresponding to the mapping feature suppression region is 0, the multiplication forces all the transmission weights of the feature extraction channel corresponding to the feature suppression region to 0. Thus, the transmission weights of the feature extraction channel are truncated to zero through the topology control mask, so that the power feature of the interfered frequency band cannot be further convolved and amplified in the subsequent residual block, blocking the extension and transmission of the feature of the interfered frequency band within the preset recognition structure.

[0035] Methods for compensating for the response gain in the clean frequency band include: Based on the multi-dimensional power grid environmental quality assessment report, the spectrum mask vector is retrieved from the running memory, and idle frequency bands with Boolean logical values ​​not equal to 1 and signal-to-noise ratio prediction values ​​higher than 30dB are searched in the spectrum mask vector; Extract the frequency ranges not covered by the feature suppression region in the multi-dimensional power grid environmental quality assessment report (e.g., the frequency ranges within 10Hz to 1kHz and 5kHz to 10kHz in the example) to obtain the high signal-to-noise ratio frequency band. Based on the mapping relationship between feature channels and frequency density in the structure configuration file of the preset recognition structure, i.e., the deep residual network, the specific convolutional kernel feature channels responsible for processing the frequency bands from 10Hz to 1kHz and from 5kHz to 10kHz are found in the preset recognition structure, i.e., the feature channels corresponding to the 1st to 4th feature bits and the 10th to 64th feature bits. The index address of the specific convolutional kernel feature channel is marked in the running memory, thereby locating the feature path corresponding to the high signal-to-noise ratio frequency band in the preset recognition structure as the clean channel to be compensated.

[0036] The number of feature bits in the topology control mask whose values ​​are forcibly modified to zero is counted to obtain the number of channels that are truncated (5 in the example). The number of channels (5 in the example) is multiplied by the inherent floating-point operation computing power consumption of a single feature extraction channel (10 MFLOPs in the example) to calculate the total amount of computing power surplus in the hardware processor due to the zero truncation of the transmission weight. Based on the redundant computing resources released by the number of channels truncated in the topology control mask, the total computing power is divided by the total number of pure channels to be compensated (for example, 59 in the example) to calculate the computing power surplus ratio allocated to each pure channel to be compensated. This computing power surplus ratio is used as the gain control value. The original scaling parameters (initial example value of scaling parameters is 1.0) inside a specific residual block in the clean channel to be compensated (i.e., the deep residual network) are read. The floating-point accumulation operator is called to accumulate and write the gain control value into the scaling parameters, so that the scaling parameters are dynamically increased from 1.0 to 1.08. The scaling parameters inside the clean channel to be compensated are dynamically adjusted using the gain control value.

[0037] The transit weights of the feature extraction channels, which are forced to 0 after channel-by-channel multiplication, are matrix-concatenated and dimension-joined with the scaling parameters of the clean channels to be compensated, which are increased to 1.08 after gain control adjustment. The transit weights and scaling parameters after being truncated to zero are then recombined as a whole. The reconstructed data is overwritten into the storage area of ​​the original weight coefficients loaded by the deep residual network during forward propagation in the preset recognition structure, thereby updating the global weight matrix of the preset recognition structure. The updated global weight matrix is ​​converted into a binary configuration file and written to the standby buffer of non-volatile memory for persistent storage. The recombined global weight matrix is ​​then solidified and output as environmental self-healing parameters. Through this environmentally self-healing parameter reconfiguration design, the system successfully constructs a dynamic defense mechanism at the software level. On the one hand, it completely silences the interfered feature channels, blocking the upward propagation of noise; on the other hand, while keeping the total hardware computing power (such as the surplus 50 MFLOPs in the example) constant, it accurately converts the redundant computing power into a dynamic response gain of 8% in the clean frequency band (the scaling parameter is increased from 1.0 to 1.08). This significantly improves the isolation of the interference frequency band while ensuring the depth of expression of features in high signal-to-noise ratio areas, making the network more sensitive to load events in harsh power grid environments.

[0038] Methods for responding to load action events triggered by the total current signal include: The environmental self-healing parameters are fully written into the running parameter register array of the preset identification structure, i.e., the deep residual network, through bus transmission, thereby completing the global weight matrix reconstruction of the preset identification structure, i.e., the deep residual network. The forward propagation pipeline listening thread of the preset identification structure, namely the deep residual network, is started, making it ready to receive and dynamically parse the streaming input power data. This loads the environmental self-healing parameters into the preset identification structure, enabling the preset identification structure to run in real-time monitoring under the configuration of the environmental self-healing parameters.

[0039] During the real-time monitoring process of the preset identification structure, namely the deep residual network, the total current signal after time-domain phase compensation contained in the original sampling time series with fixed period is continuously acquired. The root mean square calculation operator is called to calculate the effective value of the current in the current power frequency cycle in the form of a sliding window. The first-order difference operator is used to subtract the effective value of the current in the current power frequency cycle from the effective value of the current in the previous adjacent power frequency cycle to obtain the rate of change of the effective value of the current. The total current signal in the current power frequency cycle is multiplied and integrated with the collected bus voltage signal to obtain the current real-time power, and then the real-time power is subtracted from the real-time power of the previous adjacent power frequency cycle to obtain the real-time power increment. Based on the proportion of odd harmonic energy contained in the characteristic suppression zone (the numerical value of the frequency band boundary of the characteristic suppression zone is taken as 1kHz to 5kHz in the example) in the multi-dimensional power grid environmental quality assessment report, the proportion of odd harmonic energy is used as the background noise distortion rate in the running memory. The background noise distortion rate is accumulated with the constant coefficient 1 using the floating-point addition operator to obtain the correction gain coefficient containing the energy distortion ratio of the characteristic suppression zone (example value is 1.2). The original fixed action threshold (0.5A in the example) is multiplied by the correction gain coefficient (1.2 in the example) which includes the energy distortion ratio of the characteristic suppression region, to obtain the new action current threshold after adaptive background noise correction (0.6A in the example). The new action current threshold is then written into the dynamic control register as the latest dynamic action threshold, thereby establishing the dynamic action threshold based on the characteristic suppression region in the multidimensional power grid environmental quality assessment report.

[0040] Compare the rate of change of the effective current value with the dynamic action threshold (0.6A in the example), or compare the real-time power increment with the corresponding dynamic power action threshold; When the effective rate of change of current is greater than 0.6A, or the real-time power increment is greater than the dynamic power action threshold, a high-level trigger signal is output to determine that the total current signal meets the event triggering condition. Then, a load interrupt trigger command containing the absolute timestamp of the current high-frequency sampling clock source is generated. When the dynamic action threshold is exceeded, it is determined that the total current signal has triggered a load action event.

[0041] Traditional non-intrusive load monitoring systems often suffer from trajectory distortion due to high-frequency noise contamination when converting voltage-current trajectories into state matrices. Furthermore, subsequent deep learning networks rely entirely on data-driven feature extraction, lacking the constraints of underlying electrical physics mechanisms, and are prone to producing irrational identification results that violate electrical laws and energy conservation. Therefore, this invention reconstructs spatiotemporal trajectories through spatial gridding denoising and introduces Kirchhoff's laws and the energy continuity equation to physically examine intermediate features of the network, and performs closed-loop correction in conjunction with energy conservation. The specific steps are as follows: Methods for outputting the spatiotemporal trajectory load state matrix include: When a load action event is received, the transient power data stored in the cyclic streaming sliding window in the static random access memory is retrieved. Based on the absolute timestamp contained in the load interruption trigger instruction, the total current signal after time-domain phase compensation with a fixed number of cycles (5 complete power frequency cycles in the example) is extracted as the power consumption fluctuation data. Retrieve the timestamp data of the zero-crossing points and peak points of the waveform locked in the basic power consumption dataset. Retrieve the continuous discrete voltage sampling points in the static random access memory that are completely consistent with the absolute timestamps of the five complete power frequency cycles covered by the power consumption fluctuation data. Extract them as transient voltage sequences that are in the same sampling period as the power consumption fluctuation data.

[0042] A two-dimensional data matrix storage area for coordinate transformation is allocated in the running memory. The value of each discrete voltage sampling point in the transient voltage sequence with the same sampling period is retrieved as the horizontal axis parameter X of the rectangular coordinate system. The value of each current sampling point corresponding to the same timestamp in the backtracked power fluctuation data with a fixed number of cycles is retrieved as the vertical axis parameter Y of the rectangular coordinate system. Based on the geometric mapping formula between rectangular coordinates and polar coordinates, the formula for calculating the polar radius is as follows: The formula for calculating the polar angle is: ;in It is the arctangent function; The coordinate pairs (X,Y) in the Cartesian coordinate system are converted into the polar radius and polar angle in the polar coordinate system. The polar radius and polar angle values ​​corresponding to each sampling time are written into the two-dimensional data matrix storage area for time axis continuity fitting. Thus, the transient voltage sequence is used as the horizontal axis parameter and the power consumption fluctuation data is used as the vertical axis parameter to map to the polar coordinate system, generating a voltage-current trajectory curve.

[0043] Based on the frequency band boundary values ​​defined as the characteristic suppression region in the multidimensional power grid environmental quality assessment report (example values ​​are within 1kHz to 5kHz), the distortion burr texture components with frequencies within 1kHz to 5kHz in the voltage-current trajectory curve are smoothed by the Laplacian operator and contrast-enhanced denoising in the polar coordinate system to obtain the enhanced voltage-current trajectory curve. Construct a two-dimensional discrete square grid space with a preset resolution (32 x 32 in the example) in the running memory, calculate the trajectory boundary of the enhanced voltage-current trajectory curve in the polar coordinate system, and scale the trajectory boundary proportionally to each grid node in the two-dimensional discrete square grid space to complete the discrete grid space mapping. The cumulative number of times the trajectory curve passes through each grid node is counted and converted into the gray value of the corresponding pixel (in the example, the value is an integer between 0 and 255). Pixel mapping is completed and a 32x32 two-dimensional gray matrix is ​​generated. The two-dimensional gray matrix is ​​then solidified and output as a spatiotemporal trajectory load state matrix.

[0044] Methods for outputting the load identification probability matrix include: The spatiotemporal trajectory load state matrix is ​​input into a deep residual network running in real-time monitoring. Multiple consecutive residual blocks with different kernel sizes are used to perform two-dimensional sliding convolution with a stride of 1 pixel for example. The kernel weights are multiplied and accumulated with the pixel gray values ​​of the overlapping area in the spatiotemporal trajectory load state matrix, and then the bias parameter is added. After being mapped by a nonlinear activation function, it is passed backward to perform continuous matrix operations, thereby extracting features at different feature resolution scales. Among them, the example value of the shallow residual block in the front layer of the deep residual network is the residual block from the first convolutional layer to the fourth layer. The edges, stripes, corners, grid boundary lines and discrete trajectory point distribution patterns that reflect the local geometry of the voltage-current trajectory curve are extracted as shallow spatial texture features. The middle layer residual block example takes the values ​​from the fifth layer residual block to the eighth layer residual block. The middle layer abstract features reflecting the fusion of local trajectory features and the mid-dimensional transition topology are extracted by downsampling. The example of the deep residual block in the back is taken from the residual block of the ninth layer to the last fully connected layer. By continuously downsampling and accumulating the number of channels, high-dimensional abstract topological features that reflect the global envelope closure shape of the trajectory curve, the bias direction of electrical properties, and the capacitive hysteresis trend are extracted as deep semantic features. By pointing memory pointers to the activation output addresses of the front shallow residual block, the middle intermediate residual block, and the rear deep residual block respectively, the intermediate state feature matrix is ​​copied to an independent memory buffer in real time to obtain multi-scale activation feature maps output by multiple internal hidden layers.

[0045] Multiply the channel values ​​in the multi-scale activation feature map by the conductance conversion coefficient (0.01 Siemens in the example) to convert them into virtual node current vectors with the physical dimension of current. According to Kirchhoff's current law electromagnetic physics calculation formula, perform algebraic sum accumulation on the virtual node current vectors corresponding to each channel output by the same residual block, and verify whether the algebraic sum of the current at any simulation node is conserved to zero to obtain the first constraint residual. The simulation node is a virtual equivalent circuit node constructed by mapping the channel flow direction of the feature map according to the preset recognition structure. The channel values ​​in the multi-scale activation feature map are converted into virtual power components through a preset power mapping factor (the preset power mapping factor is 10 watts per unit activation value in the example). The virtual power components are integrated and accumulated on the time axis (or: the virtual node current vector is multiplied by the current corresponding transient voltage sequence sample value to obtain the single-point virtual power, and integrated in the time domain within the time interval between adjacent sampling nodes) to obtain the virtual energy distribution. According to the electromagnetic physics calculation relationship of the energy continuity equation, the time differential change of virtual energy between two adjacent sampling nodes is calculated and a preset loss dissipation value is subtracted (the example value is 5 watts). It is verified whether the deviation between the remainder after subtracting the preset loss dissipation value from the change of virtual energy between two adjacent sampling nodes and zero is within the example ±0.02. If the deviation exceeds ±0.02, it is determined that the energy conversion before and after is discontinuous, and the second constraint residual is obtained. The virtual current residuals after feature mapping of each layer are calculated by weighted summation of the first constraint residual and the second constraint residual (the weight of the first constraint residual is 0.6 and the weight of the second constraint residual is 0.4). The signal-to-noise ratio prediction value of 30dB corresponding to the current timestamp in the environmental background noise fingerprint database is retrieved and converted into a safe energy fluctuation range threshold of 0.05. When the absolute value of the virtual current residual of a certain channel at a specific sampling node is greater than the threshold of 0.05, it is determined that there is an abnormal jump in the electrical activation component in the multi-scale activation feature map.

[0046] When an abnormal jump is detected, the weight coefficient of the convolution kernel corresponding to the feature propagation path with the abnormal jump is directly multiplied by the zero value coefficient, so that the weight value corresponding to the feature propagation path is forced to zero, thereby reducing the contribution weight of the feature propagation path to a state where it can no longer make a forward propagation contribution to subsequent feature extraction. The preset recognition structure, namely the deep residual network, is invoked to skip the hidden layer with abnormal jumps. The original phase information contained in the total current signal after the phase space label has been added to the basic power dataset is retrieved. The original phase information is concatenated and recombined with the multi-scale activation feature map without abnormal jumps in the channel dimension for phase alignment correction. After the recombined composite feature matrix flows through the fully connected layer of the preset recognition structure, namely the deep residual network, and the normalized exponential function operator is called to perform probability mapping calculation, the load identification probability matrix is ​​solidified in the running memory.

[0047] Extract the identification probability values ​​calculated for each type of appliance from the load identification probability matrix, and compare the identification probability values ​​of each type of appliance with a fixed probability threshold (0.50 in the example). When the identification probability value corresponding to a certain type of electrical appliance is greater than 0.50, the electrical appliance type is extracted as a valid identification item, and the device identification information of the electrical appliance type is written into a contiguous storage list area in the running memory to generate a set of candidate load components.

[0048] Using the device identification information contained in the candidate load component set as the index key, retrieve the standard electrical power fingerprint (example value 1200 watts) and standard phase feature (example value phase angle 30 degrees) corresponding to each candidate load component from the electrical feature library. Extract the total current signal that is centrally mapped and bound to the phase space label from the basic power data set, and extract the bus voltage signal from the original sampling time series with a fixed period to calculate the current measured power and measured phase. The difference between the standard electrical power fingerprint and the measured power is calculated, and the deviation between the standard phase feature and the measured phase is checked. When both the absolute power difference and the angle deviation are less than the preset matching tolerance, that is, the absolute power difference is less than the power tolerance threshold in the preset matching tolerance (50 watts in the example), and the angle deviation is less than the phase tolerance threshold in the preset matching tolerance (3.0 degrees in the example), it is confirmed that the electrical type in the candidate load component set matches the electrical component in the current total current signal, thereby identifying the electrical type.

[0049] Methods for outputting load monitoring results include: Based on the identified electrical appliance types, the corresponding standard electrical appliance power fingerprint is retrieved (the example value of the standard electrical appliance power fingerprint is 1200 watts). Combined with the bus voltage signal extracted from the basic power data set, the current components of each identified electrical appliance type are superimposed in the time domain using the waveform reconstruction method of discrete point multiplication to synthesize a segment of the expected simulated total current waveform. The expected simulated total current waveform is integrated over the entire cycle and subjected to root mean square transformation to calculate the overall power value represented by the expected simulated total current waveform. The overall power value is then solidified and extracted as the total power of the sub-loads.

[0050] The measured power of the total current signal extracted from the basic electricity dataset is compared with the total power of the sub-loads to obtain the energy deviation residual. When the energy deviation residual is within the allowable range of the electricity metering standard, it is determined that the energy conservation constraint is met.

[0051] The measured power of the total current signal is calculated by performing a dot product and integral operation on the total current signal and the bus voltage signal in the basic power data set (example value is 1210 watts). Subtract the measured power from the total power of the sub-loads, 1200 watts, and divide by the measured power, 1210 watts, to calculate the percentage and obtain the energy deviation residual (example value is 0.83%). The energy deviation residual of 0.83% is compared with the allowable range of the electricity metering standard (2% in the example). Since 0.83% is less than 2%, the current energy deviation residual is determined to be within the allowable range of the electricity metering standard, and the energy conservation constraint is satisfied.

[0052] If the calculated energy deviation residual (example value 4.5%) is greater than 2% of the allowable range of the electricity metering standard, it is determined that the energy deviation residual exceeds the allowable range of the electricity metering standard. At this point, the parameter fine-tuning logic is triggered by the feedback command. At the output of the fully connected layer of the deep residual network, the bias parameter is adjusted by adding a correction amount proportional to the energy deviation residual of 4.5%. The expected simulated total current waveform is subtracted point by point from the total current signal in the basic power data set to generate a difference waveform. This difference waveform is then used as a new input data stream and fed back into the deep residual network running in real-time monitoring for multi-scale feature extraction and secondary feature extraction. The load identification probability matrix is ​​output again and probability mapping is calculated to correct the identification of the electrical appliance type.

[0053] When determining that the energy conservation constraint is met, based on the identified appliance type, the corresponding device name (e.g., inverter air conditioner), operating mode (e.g., heating low fan mode), and energy consumption (e.g., 1.2 kWh per hour) are extracted. Simultaneously, based on the total harmonic distortion rate recorded in the current power grid quality assessment report, the matching prompt text is retrieved from the local diagnostic knowledge base (the local diagnostic knowledge base is constructed by establishing a one-to-one key-value pair mapping relationship between historical total harmonic distortion rate and other power quality data and corresponding power consumption optimization suggestions, and persistently storing it in non-volatile memory in the form of a binary configuration file) as a power consumption health suggestion (example suggestion is to run during off-peak hours to reduce harmonic interference). The equipment name, operating mode, energy consumption, and electricity health advice are concatenated into strings and encapsulated into a matrix according to a structured text format, and the load monitoring results are output. By utilizing Laplace smoothing and gridded grayscale pixel mapping, distortion spikes in the trajectory curve within the range of 1kHz to 5kHz are completely filtered out. By introducing Kirchhoff's laws and the energy continuity equation as "physical vigilance," the system can fuse abnormal jump channels with deviations exceeding 0.05 in real time. Under the energy conservation closed-loop correction mechanism, the power deviation (such as 4.5% in the example, reduced to 0.83%) is firmly limited to the 2% measurement standard. The mechanism eliminates the phenomenon of false load and significantly improves the physical robustness and rigor of equipment multi-state identification.

[0054] In summary, this invention extracts background power quality data and generates a power grid environmental quality assessment report, accurately quantifies and determines the noise distribution characteristics and purity level of the current complex power grid environment on the user side, applies it to the monitoring front end to lock harmonic interference to provide topology control, enables deep learning networks to have hardware-level environmental adaptation and noise resistance self-healing capabilities, and improves the robustness of the algorithm in harsh power grid environments. By performing real-time physical constraint review on the activated feature map and triggering feedback adjustment and secondary feature extraction based on power deviation residuals at the output end, the probabilistic prediction of deep learning and the deterministic classical electromagnetic physics laws are deeply integrated and closed-loop verified at the feature layer and output layer. This is applied to the identification and decision-making stage to conduct compliance review and reverse self-correction of the output results to eliminate false identifications, ensuring that the diagnostic report meets the energy conservation constraint and reducing the false alarm rate. The combination of these two approaches enhances the anti-interference capability of the detection system under complex and variable power grid interference, effectively ensuring the high credibility of the appliance type and energy consumption diagnostic report at the physical reality level, and making the monitoring results closer to the actual energy conservation state.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A non-intrusive load monitoring system based on deep residual networks, characterized in that, include: The acquisition module acquires the bus voltage signal and the total current signal, performs phase anchoring on the total current signal based on the voltage zero crossing point, obtains the basic power data set, extracts background power quality data, and generates a power grid quality assessment report. The self-healing module blocks the signal transmission of the interfered frequency band in the preset identification structure based on the power grid quality assessment report, compensates for the response gain of the clean frequency band, and outputs environmental self-healing parameters. The method for compensating for the response gain of the clean frequency band includes: Extract the frequency range not covered by the feature suppression region in the multi-dimensional power grid environmental quality assessment report, obtain the high signal-to-noise ratio frequency band, and locate the feature path corresponding to the high signal-to-noise ratio frequency band in the preset identification structure as the pure channel to be compensated. The gain control quantity is obtained based on the redundant computing resources released after blocking the interfered frequency band, and the scaling parameters inside the clean channel to be compensated are dynamically adjusted using the gain control quantity. The zero-truncated propagation weights and scaling parameters are recombined as a whole, the global weight matrix of the preset recognition structure is updated, and the recombined global weight matrix is ​​solidified and output as environmental self-healing parameters. The mapping module, based on the self-healing parameters of the loaded environment, responds to the load action event triggered by the total current signal, extracts the corresponding power consumption fluctuation data, maps it to the polar coordinate system, and outputs the spatiotemporal trajectory load state matrix. The method for outputting the spatiotemporal trajectory load state matrix includes: In response to load action events, backtrack and extract power consumption fluctuation data with a fixed number of cycles, and extract transient voltage sequences with the same sampling period as the power consumption fluctuation data based on the zero crossing point and peak point of the waveform. By mapping the transient voltage sequence as the horizontal axis parameter and the power consumption fluctuation data as the vertical axis parameter to the polar coordinate system, a voltage-current trajectory curve is generated. The voltage-current trajectory curve is denoised and enhanced by using the feature suppression region in the multidimensional power grid environmental quality assessment report. The enhanced voltage-current trajectory curve is then extracted by discretized grid spatial mapping and pixel mapping to output the spatiotemporal trajectory load state matrix. The verification module performs real-time logical verification of the response state of the spatiotemporal trajectory load state matrix under electromagnetic physical constraints and outputs a load identification probability matrix. The method for identifying the output load probability matrix includes: The spatiotemporal trajectory load state matrix is ​​input into a preset recognition structure for multi-scale feature extraction, resulting in multi-scale activation feature maps output by multiple internal hidden layers. Based on Kirchhoff's current law and the energy continuity equation, a real-time electromagnetic physical constraint review is performed on the multi-scale activation feature map to determine whether there are abnormal jumps in the electrical activation components of the multi-scale activation feature map. When an abnormal jump is detected, the contribution weight of the feature propagation path is reduced to a state where it cannot contribute to the forward propagation of subsequent feature extraction. The original phase information in the basic power dataset is reorganized into the multi-scale activation feature map using the jump connection structure of the preset recognition structure for phase alignment correction, and the load identification probability matrix is ​​output. The diagnostic module identifies the appliance type based on the load identification probability matrix and outputs the load monitoring results when the measured power of the total current signal and the total power of the identified sub-loads satisfy the energy conservation constraint. The method for identifying electrical appliance types based on a load identification probability matrix includes: Extract the identification probability value corresponding to each electrical appliance type in the load identification probability matrix, and output the candidate load component set when the identification probability value is greater than a fixed probability threshold; Obtain the standard electrical power fingerprint and standard phase features corresponding to the candidate load component set in the electrical feature library; match and verify the standard electrical power fingerprint and standard phase features with the basic power dataset to identify the electrical appliance type; The method for obtaining the output load monitoring results includes: Obtain the standard electrical power fingerprint to synthesize the expected simulated total current waveform, and extract the total power of the sub-loads from the expected simulated total current waveform; The measured power of the total current signal extracted from the basic electricity dataset is compared with the total power of the sub-loads to obtain the energy deviation residual. When the energy deviation residual is within the allowable range of the electricity metering standard, it is determined that the energy conservation constraint is met. When the residual energy deviation exceeds the allowable range of the electricity metering standard, the bias parameters of the output layer of the preset identification structure are adjusted by the feedback command, and the difference waveform obtained by subtracting the expected simulated total current waveform from the total current signal point by point is fed back to the preset identification structure for secondary feature extraction to correct the identification of the electrical appliance type. While satisfying the energy conservation constraint, the system integrates and identifies the equipment name, operating mode, energy consumption, and electricity health recommendations corresponding to the type of electrical appliance, and outputs load monitoring results.

2. The non-intrusive load monitoring system based on deep residual networks as described in claim 1, characterized in that, The method for obtaining the basic electricity dataset includes: Based on the preset hardware magnetic saturation correction parameters, the total current signal is subjected to time-domain phase compensation, and the phase-compensated total current signal and the acquired bus voltage signal are synchronously buffered to extract the original sampling time sequence with a fixed period. Lock the zero-crossing point and peak point of the bus voltage signal, and extract the relative time difference between each sampling point in the total current signal and the zero-crossing point of the waveform to generate phase space labels; Phase space labels are mapped and bound to the total current signal, and the bound total current signal, bus voltage signal, and waveform zero-crossing points and peak points are combined to obtain the basic power data set.

3. The non-intrusive load monitoring system based on deep residual networks as described in claim 2, characterized in that, The method for generating a power grid quality assessment report includes: The bus voltage signal is subjected to frequency domain mapping transformation to obtain the corresponding voltage spectrum feature distribution. The total harmonic distortion rate, the proportion of odd harmonic energy and the statistical characteristics of voltage fluctuation are extracted from the voltage spectrum feature distribution and combined into background power quality data. The background power quality data is compared with the preset standard to identify the interference frequency band that exceeds the preset standard. The interference frequency band is defined as the feature suppression area. The background power quality data is entered into the real-time constructed environmental background noise fingerprint database. The signal-to-noise ratio prediction value and frequency confidence in the environmental background noise fingerprint database are extracted. The feature suppression region, signal-to-noise ratio prediction value, and frequency confidence are jointly and structurally encapsulated to generate a multi-dimensional power grid environmental quality assessment report.

4. The non-intrusive load monitoring system based on deep residual networks as described in claim 3, characterized in that, The method for blocking signal transmission in the interfered frequency band of the preset identification structure includes: The purity level of the current power grid environment is determined by the predicted signal-to-noise ratio and frequency confidence level. Based on the purity level, a topology control command is triggered to initialize a control feature vector corresponding to the number of feature channels of the preset recognition structure. Set the feature bit values ​​of the frequency band features corresponding to the feature suppression region in the control feature vector to zero, generate a topology control mask, and use the topology control mask to map the feature extraction channel corresponding to the feature suppression region in the preset recognition structure. By using a topology control mask to truncate the propagation weights of the feature extraction channel to zero, the extension and propagation of features in the interfered frequency band within the preset recognition structure is blocked.

5. The non-intrusive load monitoring system based on deep residual networks as described in claim 1, characterized in that, The method for responding to a load action event triggered by a total current signal includes: The environmental self-healing parameters are loaded into the preset identification structure, so that the preset identification structure operates in real-time monitoring state under the configuration of the environmental self-healing parameters; In real-time monitoring, the effective rate of change of the total current signal and the real-time power increment are extracted in real time, and a dynamic action threshold is established based on the characteristic suppression zone in the multi-dimensional power grid environmental quality assessment report. The effective rate of change of the current signal or the real-time power increment is compared with the dynamic action threshold, and if the dynamic action threshold is exceeded, it is determined that the total current signal has triggered a load action event.

Images