A non-intrusive air conditioning load decomposition method and storage medium

Abstract

This application discloses a non-intrusive air conditioning load decomposition method and storage medium. The method includes: acquiring a high-frequency current sampling sequence at the user's electricity meter at a sampling rate of not less than 6400Hz, and obtaining the time-domain waveform sequence and frequency-domain harmonic sequence of the high-frequency current sampling sequence; constructing parallel time-domain convolutional networks and frequency-domain convolutional networks as feature extraction networks; obtaining a trained feature extraction network through two-stage training; inputting the real-time acquired high-frequency current sampling sequence into the trained feature extraction network; and outputting a reconstructed air conditioning power waveform. This invention solves the problems of easy confusion between air conditioning load and resistive load and loss of long-period dependencies by using time-frequency dual-domain feature fusion and two-stage decoupled modeling, significantly improving the decomposition accuracy and robustness.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, specifically to a non-intrusive air conditioning load decomposition method and storage medium. Background Technology

[0002] Non-Intrusive Load Monitoring (NILM) technology analyzes aggregated signals such as total current and voltage at the user's power inlet to identify the operating status and power decomposition of individual electrical devices, and is one of the key technologies for smart electricity consumption and demand-side management.

[0003] Compared to traditional invasive monitoring, this method eliminates the need to install sensors inside every appliance, offering advantages such as low cost, easy deployment, and high user acceptance. With the development of artificial intelligence technology, deep learning-based NILM methods have become a research hotspot, widely applied in areas such as home energy efficiency analysis, load forecasting, and anomaly detection.

[0004] As a major energy-consuming device in residential and commercial buildings, air conditioners face unique technical challenges in load decomposition. First, air conditioner operation modes are complex and diverse, encompassing both microsecond-level peak currents and millisecond-level transition processes during compressor startup, as well as continuous power modulation and fluctuations during steady-state operation. Traditional single-feature models struggle to simultaneously cover these two distinct operating states. Second, a complete cooling or heating cycle of an air conditioner typically lasts from several minutes to tens of minutes, exhibiting significant long-cycle time-series dependencies. Existing NILM methods based on convolutional neural networks (CNNs) have limited receptive fields, making it difficult to capture the complete operating logic of the air conditioner within long time windows, leading to power tracking loss. Furthermore, as a high-power nonlinear load, the odd-order harmonic components in the steady-state current characteristics of air conditioners, while usable for identification, are easily confused with the steady-state characteristics of high-power resistive loads such as water heaters and electric ovens in practical applications, severely limiting decomposition accuracy.

[0005] To address the aforementioned issues, existing technologies have attempted to improve the situation by increasing the sampling frequency, introducing harmonic analysis, or employing a sequence-to-sequence (Seq2Seq) model. However, the following shortcomings remain: First, while the introduction of high-frequency data can capture transient features, it lacks effective utilization of steady-state nonlinear fingerprints. Second, although end-to-end deep learning methods can automatically extract features, they are susceptible to background noise interference and lack robustness in decomposition under low signal-to-noise ratio environments. Third, a single model cannot simultaneously handle two tasks with different characteristics: air conditioner status recognition and accurate power reconstruction. Summary of the Invention

[0006] This application provides a non-intrusive air conditioning load decomposition method and storage medium to solve the problems of easy confusion between air conditioning and resistive loads, loss of long-cycle dependencies, and low decomposition accuracy under low signal-to-noise ratio.

[0007] This application provides a non-intrusive air conditioning load decomposition method, which specifically includes a data acquisition step, a first preprocessing step, a feature extraction network construction step, a first training step, a second training step, and an air conditioning power waveform acquisition step.

[0008] The data acquisition step involves acquiring a high-frequency current sampling sequence at the user's electricity meter at a sampling rate of no less than 6400Hz. The first preprocessing step preprocesses the acquired high-frequency current sampling sequence to obtain its time-domain waveform sequence and frequency-domain harmonic sequence. The feature extraction network construction step designs a parallel time-domain convolutional network T-CNN and a frequency-domain convolutional network F-CNN to construct a feature extraction network, which includes a linear classification layer. The first training step inputs the time-domain waveform sequence and the frequency-domain harmonic sequence into the feature extraction network and performs supervised learning training on the feature extraction network using the linear classification layer and timestamped switch labels to obtain a pre-trained feature extraction network. The second training step freezes the parameters of the pre-trained feature extraction network, inputs the real-time acquired high-frequency current sampling sequence into the frozen pre-trained feature extraction network, and trains the Transformer encoder-decoder architecture of the feature extraction network to obtain a trained feature extraction network. The air conditioner power waveform acquisition step inputs the real-time acquired high-frequency current sampling sequence into the trained feature extraction network and outputs the air conditioner power waveform.

[0009] Furthermore, the first preprocessing step specifically includes a first data transformation step and a second data transformation step.

[0010] The first data transformation step is to normalize the high-frequency current sampling sequence to obtain a time-domain waveform sequence; the second data transformation step is to perform a sliding window fast Fourier transform on the high-frequency current sampling sequence to generate a frequency-domain harmonic sequence.

[0011] Furthermore, the first training step specifically includes a second preprocessing step, a feature fusion step, and a state mask acquisition step.

[0012] The second preprocessing step is based on the feature extraction network, which aligns and concatenates the transient features of the time-domain waveform sequence with the steady-state features of the frequency-domain harmonic sequence to obtain a feature tensor; the feature fusion step is to fuse the feature tensor through the bidirectional GRU layer included in the feature extraction network to obtain a first air conditioner feature vector sequence; the state mask acquisition step is based on the switch label, which inputs the first air conditioner feature vector sequence into a linear classification layer to obtain an air conditioner operating state mask.

[0013] Furthermore, the second preprocessing step specifically includes a first feature extraction step, a second feature extraction step, a feature alignment step, and a feature concatenation step.

[0014] The first feature extraction step extracts the transient features of the time-domain waveform sequence using the temporal convolutional network T-CNN; the second feature extraction step extracts the steady-state features of the frequency-domain harmonic sequence using the frequency-domain convolutional network F-CNN; the feature alignment step aligns the transient features and the steady-state features in the time dimension; the feature concatenation step concatenates the aligned transient features and the steady-state features in the feature dimension to obtain a feature tensor.

[0015] Furthermore, the second training step specifically includes an air conditioner feature acquisition step, a feature vector selection step, and an air conditioner power waveform reconstruction step.

[0016] The air conditioner feature acquisition step is used to freeze the parameters of the pre-trained feature extraction network, and input the high-frequency current sampling sequence extracted in real time into the frozen pre-trained feature extraction network to obtain a second air conditioner feature vector sequence; the feature vector filtering step is based on the air conditioner operating state mask to perform gating filtering on the second air conditioner feature vector sequence to obtain a third air conditioner feature vector sequence; the air conditioner power waveform reconstruction step is to input the third air conditioner feature vector sequence into the Transformer encoder-decoder architecture of the feature extraction network, and use the global self-attention mechanism of the Transformer encoder-decoder architecture to output the reconstructed air conditioner power waveform.

[0017] Furthermore, the temporal convolutional network T-CNN employs multi-scale convolutional kernels.

[0018] Furthermore, the multi-scale convolution kernel includes a first receptive field and a second receptive field of different sizes. The first receptive field is used to respond to the microsecond-level peak current when the air conditioner starts, and the second receptive field is used to respond to the millisecond-level transition process when the air conditioner starts.

[0019] Furthermore, the length of the third air conditioning feature vector sequence covers at least one complete air conditioning cooling or heating adjustment cycle.

[0020] Furthermore, when the state mask is 0, the reconstructed air conditioner power waveform is not output.

[0021] This application also provides a storage medium storing computer-readable instructions that, when read by at least one processor, cause the at least one processor to perform at least one step in the non-intrusive air conditioning load decomposition method.

[0022] This application provides a non-intrusive air conditioning load decomposition method and storage medium. By constructing parallel temporal and frequency-domain convolutional networks, it extracts the transient and steady-state features of the air conditioner in the temporal and frequency domains, respectively, solving the problem of easy confusion between air conditioners and resistive loads at the physical level. A bidirectional GRU layer is used to perform temporal fusion of time-frequency features, enhancing the model's ability to perceive the air conditioner's operating state. An innovative two-stage decoupled training strategy is proposed. In the first stage, a state mask is generated through state classification learning to effectively filter background noise. In the second stage, based on the frozen feature extraction network, a Transformer encoder-decoder architecture is introduced, utilizing a global self-attention mechanism to model the complete operating cycle of the air conditioner, which can last from several minutes to tens of minutes, solving the problem of long-cycle dependency loss. Features are gated and filtered using the state mask, and power reconstruction is only performed when the air conditioner is detected to be running, avoiding the introduction of false power components from background loads. This application significantly improves the accuracy and robustness of air conditioning load decomposition in low signal-to-noise ratio environments, and can accurately reconstruct the power waveforms of the entire process of air conditioner startup, steady-state operation, and shutdown. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the non-intrusive air conditioning load decomposition method described in the embodiments of this application;

[0025] Figure 2 This is a flowchart of the first preprocessing step described in the embodiments of this application;

[0026] Figure 3 This is a flowchart of the first training step described in the embodiments of this application;

[0027] Figure 4 This is a flowchart of the second preprocessing step described in the embodiments of this application;

[0028] Figure 5 This is a flowchart of the second training step described in the embodiments of this application;

[0029] Figure 6 This is a schematic diagram of the storage medium and processor described in the embodiments of this application.

[0030] Explanation of reference numerals in the attached figures:

[0031] 100 storage media, 200 processors. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] The existing non-intrusive air conditioning load decomposition technology has the following main technical problems: First, the air conditioning operation mode is complex and diverse, including transient features such as microsecond-level peak current and millisecond-level transition process when the compressor starts, as well as continuous power modulation and nonlinear harmonic features during steady-state operation. Traditional single feature extraction models are difficult to cover these two completely different operating states at the same time, resulting in insufficient feature representation capabilities.

[0034] Secondly, a complete cooling or heating cycle of an air conditioner typically lasts from several minutes to tens of minutes, exhibiting significant long-cycle time-series dependence characteristics. However, existing NILM methods based on convolutional neural networks have limited receptive fields, making it difficult to capture the complete operating logic of the air conditioner within a long time window, resulting in the loss of long-cycle power fluctuation tracking.

[0035] Furthermore, while the odd harmonic components in the steady-state current characteristics of air conditioners, as high-power nonlinear loads, can be used for identification, they are easily confused with the steady-state characteristics of high-power resistive loads such as water heaters and electric ovens in practical applications, leading to a significant decrease in decomposition accuracy. Simultaneously, traditional end-to-end power regression methods are susceptible to background noise interference; in low signal-to-noise ratio environments, the model is easily skewed by noise, failing to accurately distinguish the air conditioner's operating state from the background load, and lacking an effective noise filtering mechanism.

[0036] To address the aforementioned problems, this application includes the following technical features.

[0037] like Figure 1 As shown, this application provides a non-intrusive air conditioning load decomposition method, which specifically includes step S1) data acquisition step, step S2) first preprocessing step, step S3) feature extraction network construction step, step S4) first training step, step S5) second training step and step S6) air conditioning power waveform acquisition step.

[0038] Step S1) Data acquisition step: Acquire high-frequency current sampling sequence at the user's electricity meter at a sampling rate of not less than 6400Hz.

[0039] The technical advantage lies in its use of a high-frequency sampling rate of at least 6400Hz to collect the total current sequence at the user's electricity meter. This allows for the complete capture of the microsecond-level startup spike current and millisecond-level transition process generated during the operation of the air conditioner compressor. Simultaneously, it ensures that odd-order harmonic components (such as the 3rd and 5th harmonics) are accurately collected without aliasing, providing a high-quality raw data foundation for subsequent time-frequency dual-domain feature extraction. If the sampling rate is lower than 6400Hz, the steep pulse waveform at the moment of air conditioner startup cannot be distinguished, causing the time-domain branch to lose crucial transient features. Furthermore, according to the Nyquist sampling theorem, low-frequency sampling causes harmonic aliasing, preventing the frequency-domain branch from accurately extracting the nonlinear fingerprint of the air conditioner, and thus failing to effectively distinguish between the air conditioner and resistive loads. Therefore, this high-frequency sampling step ensures the integrity and distinguishability of both time-domain transient information and frequency-domain steady-state information from the data source, which is the physical basis for the high-precision air conditioner load decomposition achieved in this invention.

[0040] Step S2) First preprocessing step: preprocess the collected high-frequency current sampling sequence to obtain the time-domain waveform sequence and frequency-domain harmonic sequence of the high-frequency current sampling sequence.

[0041] like Figure 2 As shown, step S2) the first preprocessing step specifically includes step S21) the first data transformation step and step S22) the second data transformation step.

[0042] Step S21) First data transformation step: Normalize the high-frequency current sampling sequence to obtain a time-domain waveform sequence.

[0043] Step S22) Second data transformation step: Perform sliding window fast Fourier transform on the high-frequency current sampling sequence to generate frequency domain harmonic sequence.

[0044] The technical advantage lies in converting the same high-frequency current sequence into a normalized time-domain waveform sequence and a frequency-domain harmonic sequence, respectively, providing physically meaningful and complementary input data for the subsequent dual-current feature extraction network. Normalization eliminates the dimensional effects caused by grid voltage fluctuations or differences in sampling amplitude, enabling the time-domain convolutional network to stably learn the morphological characteristics of the air conditioner's start-up pulse and ramp-up process. A sliding window fast Fourier transform is used to generate the frequency-domain harmonic sequence, and by extracting odd-order harmonic components, the nonlinear distortion characteristics of the air conditioner during steady-state operation are explicitly encoded into a frequency-domain representation, eliminating the need for the model to implicitly learn harmonic information from the original waveform.

[0045] This method achieves physical decoupling and synchronous alignment of transient information in the time domain and steady-state information in the frequency domain. It preserves the microstructure of the current waveform and highlights the nonlinear fingerprint of the air conditioner. This lays the data foundation for the subsequent parallel processing of two types of heterogeneous features by the feature extraction network, and fundamentally solves the technical problem that single-domain features are difficult to simultaneously characterize the complex operating modes of air conditioners.

[0046] Step S3) Feature extraction network construction step: Design parallel temporal convolutional network T-CNN and frequency domain convolutional network F-CNN to construct the feature extraction network, which includes a linear classification layer.

[0047] The technical achievement lies in constructing a dedicated dual-stream feature extraction architecture for the electrical characteristics of air conditioners by designing parallel temporal and frequency-domain convolutional networks. This architecture achieves physical perception decoupling and collaborative representation of the complex operating modes of air conditioners. The temporal convolutional network focuses on processing normalized temporal waveform sequences, capturing transient features such as microsecond-level peak currents and millisecond-level transition processes at the moment of air conditioner startup through multi-scale convolutional kernels. The frequency-domain convolutional network focuses on processing frequency-domain harmonic sequences, extracting nonlinear fingerprint features such as the amplitude of odd harmonics generated during the steady-state operation of the compressor. The two networks operate independently in parallel, avoiding mutual interference or feature annihilation caused by heterogeneous features in a single network. This allows the model to simultaneously consider the two distinct states of air conditioner operation: startup transients and steady-state operation. This feature extraction network serves as a shared foundation for the subsequent two-stage training. Its output is used for both the first-stage operating state classification and the second-stage power waveform reconstruction, achieving efficient reuse of device fingerprint features. This addresses the technical challenge of traditional methods simultaneously representing the multimodal operating characteristics of air conditioners at the network structure level.

[0048] The temporal convolutional network T-CNN employs multi-scale convolutional kernels, which include a first receptive field and a second receptive field of different sizes. The first receptive field is used to respond to the microsecond-level peak current when the air conditioner starts, and the second receptive field is used to respond to the millisecond-level transition process when the air conditioner starts.

[0049] The technical advantage lies in the fact that the temporal convolutional network employs multi-scale convolutional kernels. By setting first and second receptive fields of different sizes, it achieves the collaborative capture of transient features at multiple time scales during the air conditioner startup process. Specifically, the first receptive field uses a small-sized convolutional kernel, which can accurately respond to the peak current lasting only tens to hundreds of microseconds during the instantaneous start-up of the air conditioner compressor, capturing its steep rising edge and peak shape. The second receptive field uses a large-sized convolutional kernel, which can effectively cover the startup ramp-up and transition process lasting from several milliseconds to tens of milliseconds, perceiving the trend of gradually increasing power. The parallel use of multi-scale convolutional kernels eliminates the need for the model to trade off between temporal resolution and receptive field range, avoiding the feature loss problem caused by a fixed receptive field for a single-scale convolutional kernel, and ensuring that key transient information during the air conditioner startup phase is fully preserved. This design solves the technical challenge of simultaneously capturing multiple time-scale features during the air conditioner startup process at the network structure level, providing rich startup phase feature information for subsequent state recognition and power reconstruction.

[0050] Step S4) First training step: Input the time-domain waveform sequence and the frequency-domain harmonic sequence into the feature extraction network, and perform supervised learning training on the feature extraction network through the linear classification layer and the time-stamped switch label to obtain the pre-trained feature extraction network.

[0051] Its technical effect lies in realizing supervised learning of the fingerprint of air conditioning equipment by the feature extraction network, forcing the model to learn to distinguish the air conditioning operating status from background noise.

[0052] like Figure 3 As shown, step S4) the first training step specifically includes step S41) the second preprocessing step, step S42) the feature fusion step, and step S43) the state mask acquisition step.

[0053] Step S41) Second preprocessing step: Based on the feature extraction network, align and concatenate the time-domain transient features of the time-domain waveform sequence with the frequency-domain steady-state features of the frequency-domain harmonic sequence to obtain the feature tensor.

[0054] Step S42) Feature fusion step: The feature tensor is fused through the bidirectional GRU layer included in the feature extraction network to obtain the first air conditioner feature vector sequence.

[0055] The technical effect lies in the fact that by using a bidirectional GRU layer to model the temporal context of the spliced ​​feature tensor, the features at each time step contain contextual information, thereby enhancing the model's ability to perceive the long-term operating status of the air conditioner.

[0056] Step S43) State mask acquisition step: Based on the switch label, the first air conditioner feature vector sequence is input into the linear classification layer to obtain the air conditioner operating state mask.

[0057] The technical advantage lies in the fact that, based on timestamped switch labels, the fused feature vector sequence is input into a linear classification layer for supervised learning, outputting an air conditioner operating status mask. This training process enables the feature extraction network to learn to filter background load noise, focusing only on the electrical features related to the air conditioner. The generated operating status mask provides accurate time period selection criteria for the subsequent second-stage power reconstruction, fundamentally solving the technical problem that end-to-end regression methods are easily interfered with by background noise.

[0058] like Figure 4 As shown, step S41) the second preprocessing step specifically includes step S44) the first feature extraction step, step S45) the second feature extraction step, step S46) the feature alignment step, and step S47) the feature splicing step.

[0059] Step S44) First feature extraction step: extract the temporal transient features of the temporal waveform sequence through the temporal convolutional network T-CNN.

[0060] Step S45) Second feature extraction step: Extract the frequency domain steady-state features of the frequency domain harmonic sequence through the frequency domain convolutional network F-CNN.

[0061] Step S46) Feature alignment step: Align the time-domain transient features with the frequency-domain steady-state features in the time dimension.

[0062] Step S47) Feature concatenation step: The aligned time-domain transient features and frequency-domain steady-state features are concatenated along the feature dimension to obtain the feature tensor.

[0063] The technical effect is that the transient features of startup extracted by the time-domain convolutional network and the steady-state harmonic features extracted by the frequency-domain convolutional network are spliced ​​together in the feature dimension after being aligned in the time dimension, forming a feature tensor that integrates information from both the time and frequency domains, enabling the model to simultaneously perceive the startup impact and nonlinear fingerprint of the air conditioner.

[0064] Step S5) Second training step: Freeze the parameters of the pre-trained feature extraction network, input the real-time acquired high-frequency current sampling sequence into the frozen pre-trained feature extraction network, train the Transformer encoder-decoder architecture of the feature extraction network, and obtain the trained feature extraction network.

[0065] Its technical advantage lies in achieving high-precision power waveform reconstruction under state-aware constraints. By freezing the parameters of the pre-trained feature extraction network, the air conditioner fingerprint recognition capability learned in the first stage is fully preserved, avoiding the destruction of learned device features by the power regression task.

[0066] like Figure 5 As shown, step S5) the second training step specifically includes step S51) air conditioner feature acquisition step, step S52) feature vector filtering step and step S53) air conditioner power waveform reconstruction step.

[0067] Step S51) Air Conditioner Feature Acquisition Step: Freeze the parameters of the pre-trained feature extraction network, and input the real-time extracted high-frequency current sampling sequence into the frozen pre-trained feature extraction network to obtain the second air conditioner feature vector sequence.

[0068] Step S52) Feature vector filtering step: Based on the air conditioner operating status mask, the second air conditioner feature vector sequence is gating filtered to obtain a third air conditioner feature vector sequence. The length of the third air conditioner feature vector sequence covers at least one complete air conditioner cooling or heating adjustment cycle.

[0069] Its technical effect lies in the fact that the feature sequence is gated and filtered based on the air conditioner operation status mask output in the first stage, and the feature vector of the non-air conditioner operation period is set to zero, so that the subsequent network only focuses on the effective operation period of the air conditioner, fundamentally eliminating the interference of background load noise on power reconstruction.

[0070] Step S53) Air conditioner power waveform reconstruction step: Input the third air conditioner feature vector sequence into the Transformer encoder-decoder architecture of the feature extraction network, and use the global self-attention mechanism of the Transformer encoder-decoder architecture to output the reconstructed air conditioner power waveform.

[0071] The technical advantage lies in inputting the filtered feature sequences into the Transformer encoder-decoder architecture, utilizing its global self-attention mechanism to model a complete air conditioning operation cycle lasting from several minutes to tens of minutes. This allows the model to infer the power trends during subsequent steady-state and shutdown processes based on the features at startup, solving the problem of long-cycle power fluctuation tracking loss caused by the limited receptive field in traditional convolutional networks. This two-stage decoupled training strategy significantly improves the accuracy and robustness of air conditioning load decomposition in low signal-to-noise ratio environments.

[0072] Step S6) Air conditioner power waveform acquisition step: input the real-time acquired high-frequency current sampling sequence into the trained feature extraction network and output the air conditioner power waveform; when the state mask is 0, the reconstructed air conditioner power waveform is not output.

[0073] The technical advantage lies in achieving end-to-end online decomposition and output of the air conditioner power waveform by inputting real-time high-frequency current sequences into a trained feature extraction network. This step fully inherits the state awareness capability of the first stage and the power reconstruction capability of the second stage, enabling simultaneous output of the air conditioner operating state mask and the corresponding power waveform during actual operation. Crucially, when the state mask is 0, the system actively does not output the reconstructed air conditioner power waveform. This mechanism eliminates false power components introduced by background noise or model misjudgment at the output level, ensuring that the final output power curve strictly corresponds to the actual operating period of the air conditioner. This design not only conforms to the physical law that "power is zero when the device is off," improving the interpretability and reliability of the decomposition results, but also avoids misjudging the power of resistive loads such as water heaters and electric ovens as air conditioner power. It fundamentally solves the technical problem of traditional non-intrusive monitoring methods easily generating false identification and false alarms in low signal-to-noise ratio environments, providing users with accurate and reliable air conditioner load data for energy efficiency analysis and demand response.

[0074] like Figure 6As shown, this application also provides a storage medium 100 storing computer-readable instructions, which, when read by at least one processor 200, cause at least one processor 200 to execute at least one step in the appliance activation identification method.

[0075] The advantages of this application lie in providing a non-intrusive air conditioning load decomposition method and storage medium. By constructing parallel temporal and frequency-domain convolutional networks, it extracts the transient and steady-state features of the air conditioner in the temporal and frequency domains, respectively, thus solving the problem of easy confusion between air conditioners and resistive loads at the physical level. A bidirectional GRU layer is used to perform temporal fusion of time-frequency features, enhancing the model's ability to perceive the air conditioner's operating state. An innovative two-stage decoupled training strategy is proposed. In the first stage, a state mask is generated through state classification learning to effectively filter background noise. In the second stage, based on the frozen feature extraction network, a Transformer encoder-decoder architecture is introduced, utilizing a global self-attention mechanism to model the complete operating cycle of the air conditioner, which can last from several minutes to tens of minutes, solving the problem of long-cycle dependency loss. Features are gated and filtered through the state mask, and power reconstruction is only performed when the air conditioner is detected to be running, avoiding the introduction of false power components by background loads. This application significantly improves the accuracy and robustness of air conditioning load decomposition in low signal-to-noise ratio environments, and can accurately reconstruct the power waveforms of the entire process of air conditioner startup, steady-state operation, and shutdown.

[0076] The above provides a detailed description of a non-intrusive air conditioning load decomposition method and storage medium provided by this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A non-intrusive air conditioning load decomposition method, characterized by, Specifically, the steps include the following: The data acquisition step involves acquiring a high-frequency current sampling sequence at the user's electricity meter at a sampling rate of no less than 6400Hz. The first preprocessing step involves preprocessing the acquired high-frequency current sampling sequence to obtain the time-domain waveform sequence and the frequency-domain harmonic sequence of the high-frequency current sampling sequence. The feature extraction network construction steps involve designing a parallel temporal convolutional network T-CNN and a frequency domain convolutional network F-CNN to construct the feature extraction network, which includes a linear classification layer. The first training step involves inputting the time-domain waveform sequence and the frequency-domain harmonic sequence into the feature extraction network, and then performing supervised learning training on the feature extraction network through the linear classification layer and the time-stamped switch label to obtain a pre-trained feature extraction network. The first training step specifically includes the following steps: the second preprocessing step, based on the feature extraction network, aligning and concatenating the time-domain transient features of the time-domain waveform sequence with the frequency-domain steady-state features of the frequency-domain harmonic sequence to obtain a feature tensor; The feature fusion step involves the feature extraction network including a bidirectional GRU layer, through which the feature tensors are fused to obtain a first air conditioner feature vector sequence; the state mask acquisition step involves inputting the first air conditioner feature vector sequence into a linear classification layer based on the switch label to obtain the air conditioner operating state mask. The second training step involves freezing the parameters of the pre-trained feature extraction network, inputting the real-time high-frequency current sampling sequence into the frozen pre-trained feature extraction network, and training the Transformer encoder-decoder architecture of the feature extraction network to obtain the trained feature extraction network. as well as The step of acquiring the air conditioner power waveform involves inputting the real-time acquired high-frequency current sampling sequence into a trained feature extraction network to output the air conditioner power waveform.

2. The non-intrusive air conditioning load decomposition method as described in claim 1, characterized in that, The first preprocessing step specifically includes the following steps: The first data transformation step involves normalizing the high-frequency current sampling sequence to obtain a time-domain waveform sequence; and The second data transformation step involves performing a sliding window fast Fourier transform on the high-frequency current sampling sequence to generate a frequency domain harmonic sequence.

3. The non-intrusive air conditioning load decomposition method as described in claim 1, characterized in that, The second preprocessing step specifically includes the following steps: The first feature extraction step involves extracting the temporal transient features of the temporal waveform sequence using the temporal convolutional network T-CNN. The second feature extraction step involves extracting the steady-state features of the frequency domain harmonic sequence using the frequency domain convolutional network F-CNN. The feature alignment step aligns the time-domain transient features with the frequency-domain steady-state features along the time dimension; and The feature concatenation step concatenates the aligned time-domain transient features and frequency-domain steady-state features along the feature dimension to obtain the feature tensor.

4. The non-intrusive air conditioning load decomposition method as described in claim 1, characterized in that, The second training step specifically includes the following steps: The air conditioner feature acquisition step involves freezing the parameters of the pre-trained feature extraction network and inputting the real-time extracted high-frequency current sampling sequence into the frozen pre-trained feature extraction network to obtain the second air conditioner feature vector sequence. The feature vector filtering step involves gating the second air conditioner feature vector sequence based on the air conditioner operating status mask to obtain the third air conditioner feature vector sequence. as well as The air conditioner power waveform reconstruction step involves inputting the third air conditioner feature vector sequence into the Transformer encoder-decoder architecture of the feature extraction network, and using the global self-attention mechanism of the Transformer encoder-decoder architecture to output the reconstructed air conditioner power waveform.

5. The non-intrusive air conditioning load decomposition method as described in claim 1, characterized in that, The temporal convolutional network T-CNN employs multi-scale convolutional kernels.

6. The non-intrusive air conditioning load decomposition method as described in claim 5, characterized in that, The multi-scale convolution kernel includes a first receptive field and a second receptive field of different sizes. The first receptive field is used to respond to the microsecond-level peak current when the air conditioner starts, and the second receptive field is used to respond to the millisecond-level transition process when the air conditioner starts.

7. The non-intrusive air conditioning load decomposition method as described in claim 4, characterized in that, The length of the third air conditioning feature vector sequence covers at least one complete air conditioning cooling or heating adjustment cycle.

8. The non-intrusive air conditioning load decomposition method as described in claim 4, characterized in that, When the state mask is 0, the reconstructed air conditioner power waveform is not output.

9. A storage medium storing computer-readable instructions that, when read by at least one processor, cause at least one processor to perform at least one step in the non-intrusive air conditioning load decomposition method as claimed in any one of claims 1 to 8.

Images