A Deep Learning-Based Building Air Conditioning Load Monitoring System and Method

Abstract

This application relates to the field of building air conditioning load monitoring technology, specifically disclosing a deep learning-based building air conditioning load monitoring system and method. It introduces data processing and analysis algorithms based on artificial intelligence and deep learning technologies at the back end to combine and analyze the active and reactive power time-series data of building air conditioning systems. This captures the interactive correlation and fusion feature representation information between the active and reactive power time-series characteristics of the building air conditioning systems, thereby identifying and detecting air conditioning load mode types. This effectively overcomes the shortcomings of traditional NILM methods, using a more intelligent approach to mine the time-series fusion semantics between the reactive and active power of building air conditioning systems to achieve air conditioning load mode type identification. This adapts to the changing characteristics of different types of air conditioning load modes, providing strong support for optimizing the operation of air conditioning systems and contributing to the achievement of energy conservation and emission reduction goals.

Description

Technical Field

[0001] This application relates to the field of building air conditioning load monitoring technology, and more specifically, to a building air conditioning load monitoring system and method based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the improvement of people's living standards, the electricity consumption of air conditioning in commercial buildings has increased significantly, leading to a sharp rise in electricity demand during peak hours. Therefore, effective monitoring and analysis of building air conditioning load has become increasingly important. Traditional building air conditioning load monitoring usually relies on invasive methods, that is, directly measuring various parameters by installing sensors on each air conditioning unit. This method is not only costly to deploy and difficult to construct, but also very difficult to implement in existing buildings.

[0003] In recent years, non-intrusive load monitoring (NILM) technology has gradually gained attention. NILM does not require additional hardware to be installed on each electrical appliance (such as building air conditioning). Instead, it collects data from the main power inlet and combines it with advanced algorithms to identify the operating status of different appliances. This method has advantages such as low deployment cost, high reliability, and easy scalability, making it particularly suitable for large-scale applications. However, traditional NILM methods typically rely on simple statistical characteristics or frequency domain analysis to identify the operating status of air conditioners, making it difficult to capture the complexity and diversity of air conditioning load patterns. For example, spatial load patterns can change due to seasonal climate changes, and traditional air conditioning load monitoring methods struggle to provide sufficient resolution when faced with different types of air conditioning load patterns and power variation characteristics.

[0004] Therefore, an optimized building air conditioning load monitoring solution is desired. Summary of the Invention

[0005] This application provides a building air conditioning load monitoring system and method based on deep learning, which can effectively overcome the defects of the traditional NILM method. It realizes the identification of air conditioning load mode type by mining the time-series fusion semantics between the reactive power and active power of building air conditioning in a more intelligent way, so as to adapt to the changing characteristics of different types of air conditioning load modes, providing strong support for optimizing the operation of air conditioning system and helping to achieve energy conservation and emission reduction goals.

[0006] Firstly, a deep learning-based method for monitoring building air conditioning load is provided, including:

[0007] Smart meters are used to sample the air conditioning units in the monitored building to obtain the time queues of active power and reactive power.

[0008] The singular spectrum decomposition module is used to perform time series encoding on the time queues of active power and reactive power to obtain multiple active power time series subsequences and multiple reactive power time series subsequences.

[0009] Time-series features are extracted from the multiple active power time-series subseries and the multiple reactive power time-series subseries respectively to obtain active power time-series multi-scale coding features and reactive power time-series multi-scale coding features;

[0010] The active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features are subjected to feature alignment and fusion processing based on semantic information field modulation to obtain active power-reactive power time-series fine-grained aligned fused features. This includes: calculating the semantic information field between the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain a fine-grained semantic information field between the active power and reactive power time-series coding features; and performing feature fusion based on field mapping fine-grained alignment between the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain the active power-reactive power time-series fine-grained aligned fused features.

[0011] Based on the active power-reactive power time-series fine-grained alignment and fusion characteristics, the load monitoring results are determined, and the load monitoring results are used to represent the air conditioning load mode type label.

[0012] Secondly, a building air conditioning load monitoring system based on deep learning is provided, including:

[0013] The building air conditioning power sampling module is used to sample the air conditioning of the monitored building using a smart meter to obtain the time queue of active power and the time queue of reactive power.

[0014] The power time series encoding module is used to perform time series encoding on the time queue of active power and the time queue of reactive power using the singular spectrum decomposition module to obtain multiple active power time series subsequences and multiple reactive power time series subsequences.

[0015] The power time series feature extraction module is used to extract time series features from the multiple active power time series subsequences and the multiple reactive power time series subsequences respectively to obtain active power time series multi-scale coding features and reactive power time series multi-scale coding features.

[0016] A power feature alignment and fusion processing module is used to perform feature alignment and fusion processing based on semantic information field modulation on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain active power-reactive power time-series fine-grained aligned fusion features. The power feature alignment and fusion processing module includes: a semantic information field calculation unit, used to perform semantic information field calculation on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain a fine-grained semantic information field between the active power and reactive power time-series coding features; and a fine-grained alignment and fusion processing unit, used to perform feature fusion based on field mapping fine-grained alignment on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features based on the fine-grained semantic information field between the active power and reactive power time-series coding features to obtain the active power-reactive power time-series fine-grained aligned fusion features.

[0017] The load monitoring result determination module is used to determine the load monitoring result based on the active power-reactive power time-series fine-grained alignment fusion feature, and the load monitoring result is used to represent the air conditioning load mode type label.

[0018] This application provides a deep learning-based building air conditioning load monitoring system and method. It periodically samples the active and reactive power of air conditioning units in monitored buildings using smart meters to generate time-series data. Then, it employs data processing and analysis algorithms based on artificial intelligence and deep learning to combine and analyze the time-series active and reactive power data. This captures the interactive relationships and fusion feature representations between the time-series active and reactive power characteristics of the building's air conditioning systems, enabling the identification and detection of air conditioning load mode types (such as two-state on / off, finite-state, continuously changing, and fluctuating types). This deep learning-based building air conditioning load monitoring method effectively overcomes the shortcomings of traditional NILM methods. By intelligently mining the time-series fusion semantics between the reactive and active power of building air conditioning systems, it achieves air conditioning load mode type identification, adapting to the changing characteristics of different types of air conditioning load modes. This provides strong support for optimizing the operation of air conditioning systems and helps achieve energy conservation and emission reduction goals. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings of the embodiments of this application will be briefly described below. Obviously, the drawings described below only relate to some embodiments of this application, and are not intended to limit this application.

[0020] Figure 1 This is a schematic flowchart illustrating the deep learning-based building air conditioning load monitoring method according to an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of data flow for a deep learning-based building air conditioning load monitoring method according to an embodiment of this application.

[0022] Figure 3 This is a schematic flowchart of step S4 in the deep learning-based building air conditioning load monitoring method of this application embodiment.

[0023] Figure 4 This is a schematic flowchart of step S41 in the deep learning-based building air conditioning load monitoring method of this application embodiment.

[0024] Figure 5 This is a schematic flowchart of step S42 in the deep learning-based building air conditioning load monitoring method of this application embodiment.

[0025] Figure 6 This is a schematic block diagram of a deep learning-based building air conditioning load monitoring system according to an embodiment of this application. Detailed Implementation

[0026] The technical solutions in 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 of ordinary skill in the art without creative effort are also within the scope of protection of this application.

[0027] To address the aforementioned technical issues, this application proposes a deep learning-based method for monitoring building air conditioning loads. This method periodically samples the active and reactive power of air conditioning units in the monitored building using smart meters, generating time-series data. Then, it employs artificial intelligence and deep learning-based data processing and analysis algorithms at the backend to combine and analyze the time-series active and reactive power data. This captures the interactive relationships and fusion feature representations between the time-series active and reactive power characteristics of the building's air conditioning systems, enabling the identification and detection of air conditioning load patterns (such as two-state on / off, finite-state, continuously changing, and fluctuating). This deep learning-based method effectively overcomes the shortcomings of traditional NILM methods by intelligently mining the time-series fusion semantics between the reactive and active power of the building's air conditioning systems to identify air conditioning load patterns. This allows for adaptation to the changing characteristics of different types of air conditioning load patterns, providing strong support for optimizing the operation of air conditioning systems and contributing to the achievement of energy conservation and emission reduction goals.

[0028] Specifically, such as Figure 1 and Figure 2 As shown, the deep learning-based building air conditioning load monitoring method is characterized by comprising: S1, sampling the air conditioning units in the monitored building using smart meters to obtain a time queue of active power and a time queue of reactive power; S2, using a singular spectrum decomposition module to perform time series encoding on the time queues of active power and reactive power to obtain multiple active power time-series sub-series and multiple reactive power time-series sub-series; S3, extracting time-series features from the multiple active power time-series sub-series and the multiple reactive power time-series sub-series to obtain active power time-series multi-scale encoding features and reactive power time-series multi-scale encoding features; S4, performing feature alignment and fusion processing based on semantic information field modulation on the active power time-series multi-scale encoding features and the reactive power time-series multi-scale encoding features to obtain active power-reactive power time-series fine-grained alignment and fusion features; S5, determining the load monitoring result based on the active power-reactive power time-series fine-grained alignment and fusion features, wherein the load monitoring result is used to represent the air conditioning load mode type label.

[0029] For example, in step S1, a smart meter is used to sample the air conditioning system of the monitored building to obtain a time sequence of active power and a time sequence of reactive power. It should be understood that the smart meter can collect time-series data of active and reactive power of the air conditioning system in the monitored building at high frequency intervals (e.g., every 15 seconds), which reflect the power output of the air conditioning equipment over time.

[0030] For example, in step S2, the Singular Spectrum Decomposition (SSA) module is used to perform time series encoding on the time queues of active power and reactive power to obtain multiple active power time series subsequences and multiple reactive power time series subsequences. It should be understood that SSA is a powerful time series analysis technique that can be used to decompose, predict, and reconstruct time series data. SSA does not rely on a specific data generation model and is therefore applicable to various types of time series, including nonlinear and non-stationary series. In power systems, SSA can be used to process time series data of active and reactive power to extract meaningful subsequences, helping to understand and analyze the dynamic behavior of the system in the air conditioning of monitored buildings, reducing noise interference, and providing a basis for subsequent identification and monitoring of air conditioning load patterns.

[0031] In one embodiment, the process of using the Singular Spectrum Analysis (SSA) module to time-series encode time queues of active power and reactive power to obtain multiple active power time-series subsequences and multiple reactive power time-series subsequences includes: selecting a window length L (also known as the embedding dimension), which should be large enough to capture the periodicity and trend in the time series, but not so large as to introduce excessive noise. For each time series, a trajectory matrix is ​​constructed according to the selected window length L. For example, for the active power time series {X1,X2,...,XN}, a trajectory matrix of size L×K (where K=N-L+1) is constructed, where each column of the matrix is ​​a window of the time series, and adjacent windows differ by only one time point. Singular value decomposition is performed on the constructed trajectory matrix. This step decomposes the original time series into several orthogonal components, each carrying a portion of the information from the original time series and ordered according to the proportion of variance they explain. In this way, it is possible to identify which components correspond to noise and which correspond to the main patterns of the signal (such as trends, seasonal fluctuations, etc.). Depending on the actual needs or analysis results, components corresponding to certain singular vectors can be selected for reconstruction. For example, the most important components can be selected to reconstruct the time series, thereby removing noise and extracting the main dynamic features; or fluctuations within different frequency ranges can be separated as separate time series subsequences. Finally, the data processed through the above steps can be regarded as active power time series and reactive power time series subsequences after SSA encoding. These subsequences not only reduce noise interference in the original data, but also retain important periodic and non-periodic components, which helps to further understand and analyze the operating status of building air conditioning systems.

[0032] For example, in step S3, time-series features are extracted from the multiple active power time-series subseries and the multiple reactive power time-series subseries to obtain active power time-series multi-scale coding features and reactive power time-series multi-scale coding features. Considering that the multiple time-series subseries of active and reactive power after singular spectral decomposition have temporal correlations and mutual influences in the time dimension, these correlation features are crucial for identifying complex air conditioning load patterns, as these patterns often exhibit significant time dependence and periodicity.

[0033] In one embodiment, time-series feature extraction is performed on the multiple active power time-series subseries and the multiple reactive power time-series subseries to obtain active power time-series multi-scale coding features and reactive power time-series multi-scale coding features, respectively. This includes inputting the multiple active power time-series subseries and the multiple reactive power time-series subseries into a sequence encoder based on a bidirectional LSTM model to obtain active power time-series multi-scale coding feature vectors as active power time-series multi-scale coding features and reactive power time-series multi-scale coding feature vectors as reactive power time-series multi-scale coding features. It should be understood that through the processing of the sequence encoder based on the bidirectional LSTM model, the long-term time-series multi-scale dependency relationships and feature information of the active power and reactive power of the monitored building's air conditioning system can be captured. This is very useful for understanding the long-term operating patterns in the air conditioning system, such as the impact of seasonal changes on air conditioning usage patterns, thereby helping to better describe the diversity of air conditioning load patterns, including but not limited to two-state switching, finite-state, continuously changing, and fluctuating patterns, providing a basis for subsequent air conditioning load pattern type identification.

[0034] For example, in step S4, the active power time-series multi-scale encoded features and the reactive power time-series multi-scale encoded features are subjected to feature alignment fusion processing based on semantic information field modulation to obtain active power-reactive power time-series fine-grained aligned fusion features. It should be understood that in building air conditioning systems, there are complex interactions between active power time-series semantics and reactive power time-series semantics. For example, under different operating modes, the trends of these two power variations may be both correlated and independent. Therefore, it is necessary to capture the subtle differences in these variations and their inherent connections to support accurate identification of air conditioning load patterns. However, traditional feature fusion methods may lead to information loss or inaccurate representations, especially when processing data from different sources or modalities. Furthermore, the operation of air conditioning systems is affected not only by internal mechanisms but also by external environmental factors. This means there is a significant gap between low-level perceptual features directly obtained from power data and high-level semantic concepts. At the same time, air conditioning systems in different buildings may have different configurations and usage habits. Therefore, in the technical solution of this application, the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features are further subjected to feature alignment and fusion processing based on semantic information field modulation to obtain active power-reactive power time-series fine-grained alignment and fusion features.

[0035] In one embodiment, such as Figure 3As shown, the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features are subjected to feature alignment and fusion processing based on semantic information field modulation to obtain active power-reactive power time-series fine-grained aligned fused features, including: S41, performing semantic information field calculation on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain a fine-grained semantic information field between active power-reactive power time-series coding features; S42, based on the fine-grained semantic information field between active power-reactive power time-series coding features, performing feature fusion based on field mapping fine-grained alignment on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain the active power-reactive power time-series fine-grained aligned fused features.

[0036] In one embodiment, such as Figure 4 As shown, semantic information field calculation is performed on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain a fine-grained semantic information field between the active power and reactive power time-series coding features, including: S411, performing dimension modulation on the active power time-series multi-scale coding feature vector and the reactive power time-series multi-scale coding feature vector to obtain modulated active power time-series multi-scale coding feature vector and modulated reactive power time-series multi-scale coding feature vector; S412, performing fine-grained correlation coding and semantic information field coding on the modulated active power time-series multi-scale coding feature vector and the modulated reactive power time-series multi-scale coding feature vector to obtain a fine-grained semantic information field between the active power and reactive power time-series coding features.

[0037] In one embodiment, step S411, performing dimension modulation on the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector to obtain modulated active power time-series multi-scale encoded feature vector and modulated reactive power time-series multi-scale encoded feature vector, includes: inputting the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector into a point convolution-based dimension modulation module to obtain the modulated active power time-series multi-scale encoded feature vector and the modulated reactive power time-series multi-scale encoded feature vector, wherein the modulated active power time-series multi-scale encoded feature vector and the modulated reactive power time-series multi-scale encoded feature vector have the same feature dimension. Specifically, this process can be expressed by the formula:

[0038] v ′1 =LeakyReLU{Conv 1×1 (v1)}

[0039] v ′2 =LeakyReLU{Conv 1×1(v2)}

[0040] Wherein, v1 and v2 are the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector, respectively, Conv 1×1 (·) represents pointwise convolution, and Leaky ReLU(·) is the Leaky ReLU activation function. ′1 and v ′2 These are the time-series multi-scale encoded feature vectors of modulated active power and modulated reactive power, respectively.

[0041] It is understandable that, since the time-series data of active and reactive power may originate from different physical processes, the feature vectors generated after encoding by their respective bidirectional LSTM models may have different dimensions or structures. Directly fusing these feature vectors may lead to information loss or inaccurate representation. Pointwise convolution can adjust the dimensions of these two feature vectors, allowing them to be compared and combined in the same space during subsequent processing, thus ensuring comparability and consistency between features. Pointwise convolution, also known as 1x1 convolution, is used here to perform a linear transformation that can change the number of channels in the input feature map without affecting its spatial resolution. In this scenario, pointwise convolution is used to remap the encoded feature vectors of active and reactive power into a new feature space, where each dimension represents a new, higher-level abstract feature. This transformation helps capture more complex interactions between the original features, rather than simply superimposing or splicing them. After pointwise convolution, the generated modulated feature vectors not only retain the original time dependence and periodicity information but also introduce cross-domain information interaction, namely the interaction between active and reactive power. This helps reveal the potential correlation between the two, such as the phenomenon that the changing trends of these two powers are both correlated and independent under different operating modes. Therefore, the modulated feature vector can better reflect the actual operating state and dynamic characteristics of the air conditioning system, providing a richer and more accurate feature representation for subsequent load pattern recognition.

[0042] In one embodiment, in step S412, fine-grained correlation coding and semantic information field coding are performed on the modulated active power time-series multi-scale coded feature vector and the modulated reactive power time-series multi-scale coded feature vector to obtain the fine-grained semantic information field between the active power and reactive power time-series coded features. This includes: performing fine-grained correlation coding on the modulated active power time-series multi-scale coded feature vector and the modulated reactive power time-series multi-scale coded feature vector to obtain the active power-reactive power time-series fine-grained correlation feature matrix; and inputting the active power-reactive power time-series fine-grained correlation feature matrix into a kernel-based semantic information field coding network to obtain the fine-grained semantic information field between the active power and reactive power time-series coded features. Specifically, this process can be expressed by the following formula:

[0043]

[0044] Among them, v ′1 and v ′2 These are the modulated active power time-series multi-scale encoded feature vectors and the modulated reactive power time-series multi-scale encoded feature vectors, respectively, where L is the length of the modulated reactive power time-series multi-scale encoded feature vector, and M is the length of the modulated reactive power time-series multi-scale encoded feature vector. x The active power-reactive power time-series fine-grained correlation feature matrix, Conv 3×3 (·) represents a convolutional layer with a 3×3 kernel, and Ω represents the fine-grained semantic information field between active power and reactive power temporal coding features. This refers to matrix multiplication.

[0045] In other words, through fine-grained correlation coding, the system can capture the local dependencies of modulated active and reactive power features at various time points, revealing short-term behavioral patterns in the air conditioning system, such as start-up, shutdown, or other transient events. Simultaneously, the generated active-reactive power time-series fine-grained correlation feature matrix not only retains the time-series attributes of the original features but also introduces cross-domain information interaction, enhancing feature expressiveness. This enhanced feature representation can better reflect the dynamic behavior of the air conditioning system, including but not limited to two-state switching, finite-state, continuously changing, and fluctuating load patterns. Furthermore, fine-grained correlation coding ensures that features extracted even in different time-series contexts can accurately correspond, improving the quality and expressiveness of the fused feature representation and helping to more accurately describe the actual operating state of the air conditioning system.

[0046] Next, this fine-grained associated feature matrix is ​​input into a semantic information field encoding network based on convolutional kernels. By performing convolution operations on the feature matrix, the semantic relationships and hidden connections between the two features can be captured at a higher level. Convolution operations help the model learn the spatial structure and patterns between features, constructing a space rich in semantic information. This enables the model to understand the true meaning of the data more deeply, such as distinguishing similar but different load patterns (e.g., two-state switching versus fluctuating). The learning of convolutional layers allows the model to automatically discover and reinforce subtle semantic relationships between active and reactive power features, relationships that may not be obvious to human analysts. Furthermore, the constructed semantic information field helps mitigate domain differences, promotes effective knowledge transfer and rapid model adjustment. Even when faced with air conditioning systems with different configurations and usage habits in different buildings, the trained model can adapt well to new environments and provide accurate load pattern recognition results. Ultimately, a highly comprehensive and semantically rich temporal fine-grained aligned fusion feature representation of active and reactive power is obtained.

[0047] In one embodiment, such as Figure 5 As shown, based on the fine-grained semantic information field between the active power and reactive power time-series coding features, the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features are fused using field mapping-based fine-grained alignment to obtain the active power-reactive power time-series fine-grained aligned fused features. This includes: S421, mapping the modulated active power time-series multi-scale coding feature vector and the modulated reactive power time-series multi-scale coding feature vector to the fine-grained semantic information field between the active power and reactive power time-series coding features to obtain fine-grained aligned active power time-series multi-scale coding feature vector and fine-grained aligned reactive power time-series multi-scale coding feature vector, respectively; S422, calculating the position-weighted sum between the fine-grained aligned active power time-series multi-scale coding feature vector and the fine-grained aligned reactive power time-series multi-scale coding feature vector to obtain the active power-reactive power time-series fine-grained aligned fused feature vector as the active power-reactive power time-series fine-grained aligned fused features.

[0048] In one embodiment, in step S421, mapping the modulated active power time-series multi-scale coded feature vector and the modulated reactive power time-series multi-scale coded feature vector to the fine-grained semantic information field between active power and reactive power time-series coded features to obtain fine-grained aligned active power time-series multi-scale coded feature vector and fine-grained aligned reactive power time-series multi-scale coded feature vector includes: multiplying the modulated active power time-series multi-scale coded feature vector and the modulated reactive power time-series multi-scale coded feature vector with the fine-grained semantic information field between active power and reactive power time-series coded features to obtain fine-grained aligned active power time-series multi-scale coded feature vector and fine-grained aligned reactive power time-series multi-scale coded feature vector, respectively. Specifically, this process can be expressed by the following formula:

[0049]

[0050]

[0051] Among them, v ′1 and v ′2 These are the modulated active power time-series multi-scale encoded feature vectors and the modulated reactive power time-series multi-scale encoded feature vectors, respectively, and Ω is the fine-grained semantic information field between the active power and reactive power time-series encoded features. For matrix multiplication, v 1t and v 2t These are the fine-grained aligned active power time-series multi-scale encoded feature vectors and the fine-grained aligned reactive power time-series multi-scale encoded feature vectors, respectively.

[0052] In other words, this application utilizes a fine-grained semantic information field matrix as a medium, which contains deep-level interaction relationships extracted from the active power-reactive power time-series fine-grained correlation feature matrix through convolution operations. When this application multiplies the modulated active power and reactive power feature vectors with this semantic information field matrix, it effectively maps the original features to a new, semantically rich space. This mapping allows the active power and reactive power features at each time point to find their optimal matching position in the semantic space, thereby achieving fine-grained alignment between features. Through such alignment, even time-series features extracted in different contexts (e.g., due to changes in the air conditioning system's operating mode or the influence of external environmental factors) can achieve semantic consistency and correspondence. This not only improves the quality of feature representation but also enhances the model's ability to understand complex dynamic behaviors. For example, during the start-up, shutdown, and other transient events of an air conditioning system, the trends of active and reactive power changes may be both correlated and independent; after alignment processing, these subtle differences can be captured more accurately, which in turn helps to more finely distinguish different load mode types, such as on / off two-state, finite state, continuously changing, and fluctuating types.

[0053] In one embodiment, in step S422, a position-weighted sum is calculated between the fine-grained aligned active power time-series multi-scale encoded feature vector and the fine-grained aligned reactive power time-series multi-scale encoded feature vector to obtain an active power-reactive power time-series fine-grained aligned fusion feature vector as the active power-reactive power time-series fine-grained aligned fusion feature. Specifically, this process can be expressed by the following formula:

[0054] v c =αv 1t +βv 2t

[0055] Among them, v 1t and v 2t These are the fine-grained aligned active power time-series multi-scale encoded feature vectors and the fine-grained aligned reactive power time-series multi-scale encoded feature vectors, respectively. α and β are weight hyperparameters, and v... c The active power-reactive power time-series fine-grained aligned and fused feature vector is the active power-reactive power time-series feature vector.

[0056] In other words, this application introduces a weighting mechanism to fuse the already finely aligned active and reactive power feature vectors. The weight at each location reflects the importance or contribution of the two power features at that location, allowing the model to dynamically adjust the balance between the two based on actual conditions. For example, in some cases, changes in active power may reflect the operating status of the air conditioning system better than reactive power; while in other scenarios, the opposite may be true. Therefore, the location-weighted summation process can flexibly capture such changes, ensuring that the final fused features retain the original information while highlighting key features. The feature vector fused in this way—the active power-reactive power time-series fine-grained aligned fused feature vector—not only integrates the time-series characteristics of active and reactive power but also incorporates the complex interactions between them. This fused feature is crucial for describing the true operating state of the air conditioning system because it helps to better understand the actual operating modes and dynamic characteristics of the equipment. For example, it can more accurately distinguish between similar but different load patterns (such as on / off two-state and fluctuating types) and provide richer and more interpretable input data for subsequent classification tasks.

[0057] In summary, feature alignment and fusion processing based on semantic information field modulation enables fine-grained alignment of the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector. This ensures that features extracted even in different time-series contexts can accurately correspond, improving the quality and expressive power of the fused active power-reactive power time-series feature representation. This is crucial for describing the dynamic behavior of air conditioning systems, as it helps to better understand the actual operating state of the equipment. Furthermore, feature alignment and fusion processing based on semantic information field modulation can capture the semantic relationships and hidden connections between the active power time-series multi-scale encoded features and the reactive power time-series multi-scale encoded features at a higher level, allowing the model to understand the true meaning of the data more deeply, such as distinguishing between similar but different load patterns (e.g., two-state switching and fluctuating loads). Furthermore, by constructing a semantically rich space to achieve fine-grained alignment and deep interaction between the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector, the provided semantic information field helps alleviate domain differences, promotes effective knowledge transfer and rapid model adjustment, thereby capturing and strengthening the subtle semantic relationships between the active power time-series multi-scale encoded features and the reactive power time-series multi-scale encoded features. This generates a highly comprehensive and semantically rich active power-reactive power time-series fine-grained aligned fusion feature vector as the active power-reactive power time-series fine-grained aligned fusion feature representation, providing a basis for subsequent building air conditioning load mode type identification and classification.

[0058] For example, in step S5, the load monitoring result is determined based on the active power-reactive power time-series fine-grained alignment fusion feature, and the load monitoring result is used to represent the air conditioning load mode type label. In one embodiment, determining the load monitoring result based on the active power-reactive power time-series fine-grained alignment fusion feature, and the load monitoring result being used to represent the air conditioning load mode type label, includes: inputting the active power-reactive power time-series fine-grained alignment fusion feature vector into a classifier-based load monitoring module to obtain the load monitoring result. That is, the fine-grained feature alignment fusion representation information between the active power time-series multi-scale correlation features and reactive power time-series multi-scale correlation features of the monitored building's air conditioning is used for classification processing, thereby identifying and monitoring the building's air conditioning load mode type. It is worth mentioning that the results of building air conditioning load monitoring reveal the diversity of air conditioning load modes, which can be divided into four categories according to their power change characteristics: on / off two-state type, finite state type, continuous change type, and fluctuating type. This deep learning-based building air conditioning load monitoring method can effectively overcome the shortcomings of the traditional NILM method. It can identify the air conditioning load mode type by mining the time-series fusion semantics between the reactive power and active power of the building air conditioning in a more intelligent way, so as to adapt to the changing characteristics of different types of air conditioning load modes. This provides strong support for optimizing the operation of the air conditioning system and helps to achieve the goals of energy conservation and emission reduction.

[0059] In one embodiment, the classifier-based load monitoring module includes a fully connected (Dense Layer) and a Softmax layer. The Dense Layer maps the active power-reactive power time-series fine-grained aligned fusion feature vector to a fixed-dimensional space, preparing for subsequent classification. A Softmax layer is added after the Dense Layer. The Softmax function transforms any real value output from the previous layer into a probability distribution, where each value represents the probability that a sample belongs to the corresponding class. For a K-class problem, the Softmax layer produces a probability vector of length K, where the sum of all elements equals 1.

[0060] Here, when the active power time-series multi-scale encoded feature vector and the reactive power time-series multi-scale encoded feature vector represent the time-series multi-scale periodic encoded features of the active power and reactive power of the monitored building's air conditioning, respectively, after performing feature alignment interaction based on the semantic information field, the active power-reactive power time-series fine-grained alignment fusion feature vector will also have differences in the alignment interaction distribution space structure due to the large difference in the time-series multi-scale correspondence offset of the source data, affecting the convergence consistency of the classifier, and thus affecting the accuracy of the load monitoring results obtained by the load monitoring module.

[0061] Therefore, in one example, the active power-reactive power time-series fine-grained aligned fused feature vector is optimized. The optimization process includes:

[0062] The sum of the absolute values ​​of all eigenvalues ​​of the active power-reactive power time-series fine-grained alignment and fusion eigenvectors and the square root of the sum of their squares are used to obtain the first active power-reactive power time-series fine-grained alignment and fusion spatial structure value and the second active power-reactive power time-series fine-grained alignment and fusion spatial structure value, i.e.:

[0063]

[0064] Among them, v i w1 represents the i-th feature value of the active power-reactive power time series fine-grained alignment and fusion feature vector, w2 represents the first active power-reactive power time series fine-grained alignment and fusion spatial structure value, and w2 represents the second active power-reactive power time series fine-grained alignment and fusion spatial structure value.

[0065] Determine the number n of all eigenvalues ​​of the active power-reactive power time series fine-grained aligned fused feature vector, which is the length of the active power-reactive power time series fine-grained aligned fused feature vector;

[0066] For each eigenvalue of the active power-reactive power time-series fine-grained aligned fusion feature vector, calculate the first active power-reactive power time-series fine-grained aligned fusion long-range dependency value x by subtracting the product of the eigenvalue and the number of eigenvalues ​​from the first active power-reactive power time-series fine-grained aligned fusion spatial structure value. i =w1-v i ×n, where w1 represents the first active power-reactive power time-series fine-grained aligned fused spatial structure value, v i Let x represent the i-th eigenvalue of the active power-reactive power time-series fine-grained aligned fused feature vector, and n represent the number of eigenvalues ​​of all eigenvalues ​​in the active power-reactive power time-series fine-grained aligned fused feature vector. i This represents the first active power - reactive power time-series fine-grained alignment and fusion long-range dependency value;

[0067] The second active power-reactive power time series fine-grained alignment fusion long-range dependency value is obtained by multiplying the square root of the number of eigenvalues ​​by the product of the eigenvalues ​​and the second active power-reactive power time series fine-grained alignment fusion spatial structure value. Where w2 represents the second active power-reactive power time-series fine-grained alignment and fusion spatial structure value, v iLet y represent the i-th eigenvalue of the active power-reactive power time-series fine-grained aligned fused feature vector, and n represent the number of eigenvalues ​​of all eigenvalues ​​in the active power-reactive power time-series fine-grained aligned fused feature vector. i This represents the long-range dependency value of the second active power-reactive power time series fine-grained alignment fusion.

[0068] The exponent value calculated using the exponent of the first active power-reactive power time series fine-grained alignment and fusion long-range dependency value as a natural constant is weighted and summed with the reciprocal of the second active power-reactive power time series fine-grained alignment and fusion long-range dependency value to obtain the optimized eigenvalue corresponding to each eigenvalue. , where x i This represents the first active power - reactive power time-series fine-grained aligned and fused long-range dependency value, y. i The second active power-reactive power time series fine-grained alignment fusion long-range dependency value is represented by e, which represents the natural constant, and γ and δ represent the weighted hyperparameters. ′i This represents the optimized feature value corresponding to each feature value;

[0069] The optimized feature values ​​are used to form an optimized active power-reactive power time-series fine-grained aligned and fused feature vector.

[0070] In other words, to address the potential spatial structure deficiencies in the feature set of the active power-reactive power time-series fine-grained aligned fusion feature vector in high-dimensional space, which could lead to inconsistent convergence due to the classifier's weights implicitly inferring spatial structure information based on features, long-distance feature dependencies are established by representing the spatial structure of the active power-reactive power time-series fine-grained aligned fusion feature vector relative to its overall feature scale. This establishes the feature local connectivity of the active power-reactive power time-series fine-grained aligned fusion feature vector. Furthermore, by predicting unstructured feature points of the active power-reactive power time-series fine-grained aligned fusion feature vector, spatial ambiguity information of object feature values ​​is captured. This improves the spatial inductive bias perception capability of the feature set of the active power-reactive power time-series fine-grained aligned fusion feature vector, improves the convergence consistency of the classifier, and enhances the accuracy of load monitoring results obtained by the load monitoring module based on the classifier using the active power-reactive power time-series fine-grained aligned fusion feature vector.

[0071] In summary, the deep learning-based building air conditioning load monitoring method according to the embodiments of this application has been clarified. This deep learning-based building air conditioning load monitoring method can effectively overcome the defects of the traditional NILM method. It realizes the identification of air conditioning load mode type by mining the time-series fusion semantics between the reactive power and active power of the building air conditioning in a more intelligent way, so as to adapt to the changing characteristics of different types of air conditioning load modes, providing strong support for optimizing the operation of the air conditioning system and helping to achieve the goal of energy conservation and emission reduction.

[0072] Figure 6 This is a schematic block diagram of a deep learning-based building air conditioning load monitoring system according to an embodiment of this application. Figure 6 As shown, the deep learning-based building air conditioning load monitoring system 100 includes: a building air conditioning power sampling module 110, used to sample the monitored building air conditioning using smart meters to obtain a time queue of active power and a time queue of reactive power; a power time series encoding module 120, used to perform time series encoding on the time queues of active power and reactive power using a singular spectrum decomposition module to obtain multiple active power time series subsequences and multiple reactive power time series subsequences; and a power time series feature extraction module 130, used to extract features from the multiple active power time series subsequences and the multiple reactive power time series subsequences respectively. The sequence subsequence is subjected to time-series feature extraction to obtain active power time-series multi-scale coding features and reactive power time-series multi-scale coding features; the power feature alignment and fusion processing module 140 is used to perform feature alignment and fusion processing based on semantic information field modulation on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain active power-reactive power time-series fine-grained alignment and fusion features; the load monitoring result determination module 150 is used to determine the load monitoring result based on the active power-reactive power time-series fine-grained alignment and fusion features, and the load monitoring result is used to represent the air conditioning load mode type label.

[0073] In one embodiment, the power feature alignment and fusion processing module includes: a semantic information field calculation unit, used to perform semantic information field calculation on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features to obtain a fine-grained semantic information field between active power and reactive power time-series coding features; and a fine-grained alignment and fusion processing unit, used to perform feature fusion based on field mapping fine-grained alignment on the active power time-series multi-scale coding features and the reactive power time-series multi-scale coding features based on the fine-grained semantic information field between active power and reactive power time-series coding features to obtain the active power-reactive power time-series fine-grained aligned fused features.

[0074] In one embodiment, the power time series feature extraction module is used to: input the plurality of active power time series subsequences and the plurality of reactive power time series subsequences into a sequence encoder based on a bidirectional LSTM model to obtain active power time series multi-scale coding feature vectors as active power time series multi-scale coding features and reactive power time series multi-scale coding feature vectors as reactive power time series multi-scale coding features.

[0075] Here, those skilled in the art will understand that the specific operations of each module and unit in the aforementioned deep learning-based building air conditioning load monitoring system have been referenced above. Figures 1 to 6 The description of the deep learning-based building air conditioning load monitoring method is detailed here, and therefore, its repeated description will be omitted.

[0076] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0077] It should be understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments of this application, and are not intended to limit the scope of the embodiments of this application.

[0078] It should also be understood that, in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0079] It should also be understood that the various implementation methods described in this specification can be implemented individually or in combination, and the embodiments of this application are not limited in this respect.

[0080] Unless otherwise stated, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0082] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0083] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0084] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for monitoring building air conditioning load based on deep learning, characterized in that, The method comprises the following steps: sampling the air conditioner in the monitored building by using the smart meter to obtain the time queue of active power and the time queue of reactive power; using a singular spectrum decomposition module to perform time series coding on the time queue of active power and the time queue of reactive power to obtain a plurality of active power time sequence subsequences and a plurality of reactive power time sequence subsequences; performing time sequence feature extraction on the plurality of active power time sequence subsequences and the plurality of reactive power time sequence subsequences respectively to obtain active power time sequence multi-scale coding features and reactive power time sequence multi-scale coding features; performing feature alignment fusion processing based on semantic information field modulation on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain active power-reactive power time sequence fine-grained alignment fusion features, including: performing semantic information field calculation on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain active power-reactive power time sequence coding feature inter fine-grained semantic information field; based on the active power-reactive power time sequence coding feature inter fine-grained semantic information field, performing feature fusion based on field domain mapping fine-grained alignment on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain the active power-reactive power time sequence fine-grained alignment fusion features; based on the active power-reactive power time sequence fine-grained alignment fusion features, determining a load monitoring result, the load monitoring result being used to represent an air conditioner load mode type label; performing semantic information field calculation on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain active power-reactive power time sequence coding feature inter fine-grained semantic information field, including: performing dimension modulation on the active power time sequence multi-scale coding feature vector and the reactive power time sequence multi-scale coding feature vector to obtain a modulated active power time sequence multi-scale coding feature vector and a modulated reactive power time sequence multi-scale coding feature vector; performing fine-grained correlation coding and semantic information field coding on the modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector to obtain the active power-reactive power time sequence coding feature inter fine-grained semantic information field. 2.The deep learning-based building air conditioning load monitoring method of claim 1, wherein, performing time sequence feature extraction on the plurality of active power time sequence subsequences and the plurality of reactive power time sequence subsequences respectively to obtain active power time sequence multi-scale coding features and reactive power time sequence multi-scale coding features, including: inputting the plurality of active power time sequence subsequences and the plurality of reactive power time sequence subsequences into a sequence encoder based on a bidirectional LSTM model respectively to obtain an active power time sequence multi-scale coding feature vector as the active power time sequence multi-scale coding features and a reactive power time sequence multi-scale coding feature vector as the reactive power time sequence multi-scale coding features. 3.The deep learning-based building air conditioning load monitoring method of claim 2, wherein, The active power time sequence multi-scale coding feature vector and the reactive power time sequence multi-scale coding feature vector are subjected to dimension modulation to obtain a modulated active power time sequence multi-scale coding feature vector and a modulated reactive power time sequence multi-scale coding feature vector, comprising: The active power time sequence multi-scale coding feature vector and the reactive power time sequence multi-scale coding feature vector are input into a point convolution-based dimension modulation module to obtain the modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector, wherein the modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector have the same feature dimension. 4.The deep learning-based building air conditioning load monitoring method of claim 3, wherein, The modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector are subjected to fine-grained correlation coding and semantic information field coding to obtain an active power-reactive power time sequence coding inter-feature fine-grained semantic information field, comprising: The modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector are subjected to fine-grained correlation coding to obtain an active power-reactive power time sequence fine-grained correlation feature matrix; The active power-reactive power time sequence fine-grained correlation feature matrix is input into a convolution kernel-based semantic information field coding network to obtain the active power-reactive power time sequence coding inter-feature fine-grained semantic information field. 5.The deep learning-based building air conditioning load monitoring method of claim 4, wherein, Based on the active power-reactive power time sequence coding inter-feature fine-grained semantic information field, the active power time sequence multi-scale coding feature and the reactive power time sequence multi-scale coding feature are subjected to field domain mapping fine-grained alignment-based feature fusion to obtain an active power-reactive power time sequence fine-grained alignment fusion feature, comprising: The modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector are respectively mapped to the active power-reactive power time sequence coding inter-feature fine-grained semantic information field to obtain a fine-grained alignment active power time sequence multi-scale coding feature vector and a fine-grained alignment reactive power time sequence multi-scale coding feature vector; A position-weighted sum between the fine-grained alignment active power time sequence multi-scale coding feature vector and the fine-grained alignment reactive power time sequence multi-scale coding feature vector is calculated to obtain an active power-reactive power time sequence fine-grained alignment fusion feature vector as the active power-reactive power time sequence fine-grained alignment fusion feature. 6.The deep learning-based building air conditioning load monitoring method of claim 5, wherein, The modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector are respectively mapped to the active power-reactive power time sequence coding inter-feature fine-grained semantic information field to obtain a fine-grained aligned active power time sequence multi-scale coding feature vector and a fine-grained aligned reactive power time sequence multi-scale coding feature vector, including: multiplying the modulated active power time sequence multi-scale coding feature vector and the modulated reactive power time sequence multi-scale coding feature vector by the active power-reactive power time sequence coding inter-feature fine-grained semantic information field to obtain the fine-grained aligned active power time sequence multi-scale coding feature vector and the fine-grained aligned reactive power time sequence multi-scale coding feature vector. 7.The deep learning-based building air conditioning load monitoring method of claim 6, wherein, Based on the active power-reactive power time sequence fine-grained aligned fusion feature, a load monitoring result is determined, which is used to represent an air conditioner load mode type label, including: inputting the active power-reactive power time sequence fine-grained aligned fusion feature vector into a classifier-based load monitoring module to obtain the load monitoring result.

8. A deep learning based building air conditioning load monitoring system for performing the method of any one of claims 1 to 7, characterized in that, It includes: A building air conditioner power sampling module is configured to sample a monitored building air conditioner by using a smart meter to obtain a time queue of active power and a time queue of reactive power. A power time sequence coding module is configured to use a singular spectrum decomposition module to perform time sequence coding on the time queue of active power and the time queue of reactive power to obtain a plurality of active power time sequence subsequences and a plurality of reactive power time sequence subsequences. A power time sequence feature extraction module is configured to respectively perform time sequence feature extraction on the plurality of active power time sequence subsequences and the plurality of reactive power time sequence subsequences to obtain active power time sequence multi-scale coding features and reactive power time sequence multi-scale coding features. A power feature alignment fusion processing module is configured to perform semantic information field modulation-based feature alignment fusion processing on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain active power-reactive power time sequence fine-grained aligned fusion features, wherein the power feature alignment fusion processing module includes: a semantic information field calculation unit configured to perform semantic information field calculation on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features to obtain an active power-reactive power time sequence coding inter-feature fine-grained semantic information field; and a fine-grained alignment fusion processing unit configured to perform field domain mapping-based fine-grained alignment-based feature fusion on the active power time sequence multi-scale coding features and the reactive power time sequence multi-scale coding features based on the active power-reactive power time sequence coding inter-feature fine-grained semantic information field to obtain the active power-reactive power time sequence fine-grained aligned fusion features. A load monitoring result determination module is configured to determine a load monitoring result based on the active power-reactive power time sequence fine-grained aligned fusion features, and the load monitoring result is used to represent an air conditioner load mode type label. 9.The deep learning-based building air conditioning load monitoring system of claim 8, wherein, The power time sequence feature extraction module is configured to input the plurality of active power time sequence subsequences and the plurality of reactive power time sequence subsequences into a sequence encoder based on a bidirectional LSTM model respectively to obtain an active power time sequence multi-scale coding feature vector as the active power time sequence multi-scale coding feature and a reactive power time sequence multi-scale coding feature vector as the reactive power time sequence multi-scale coding feature.

Images