Non-intrusive load disaggregation model and method oriented to spatio-temporal feature fusion attention

Abstract

This invention discloses a non-intrusive load decomposition model and method based on spatiotemporal feature fusion attention. It extracts temporal features from the load input sequence through a temporal feature extraction module and spatial features through a spatial feature extraction module. The temporal and spatial features are concatenated and input into a convolutional attention module for weighted learning and fusion of spatiotemporal features. Finally, the fused features are used to calculate the power consumption of each load through a fully connected layer. This invention combines temporal and spatial features, performing spatiotemporal feature fusion learning through a convolutional attention module that combines temporal attention and channel attention. This facilitates effective feature extraction and reduces the impact or obscuring of features by high-power appliances.

Description

Technical Field

[0001] This invention belongs to the field of load monitoring technology, specifically relating to a non-intrusive load decomposition model and method based on spatiotemporal feature-based attention fusion. Background Technology

[0002] Load monitoring, as a visualization technology for electricity consumption, allows users to understand the energy consumption of home appliances more clearly and accurately, thus optimizing electricity management. Load monitoring can be divided into intrusive load monitoring and non-intrusive load monitoring (NILM). Due to the relatively low cost and limited scalability of intrusive load monitoring, NILM has seen significant development in recent years. NILM, when not performing event detection, can also be called load decomposition; it is a method of obtaining the power consumption of individual loads by analyzing the total power consumption.

[0003] For the NILM problem, most household appliances are often used together, leading to a significant mixing of their characteristics. This is especially true for low-power appliances, whose features are easily influenced or masked by high-power appliances, making feature extraction difficult. To more effectively highlight the importance of certain features, some researchers have recently introduced attention mechanisms into load decomposition. However, existing literature mostly uses single spatial or temporal features for load decomposition. Furthermore, for most convolutional operations, once the parameters are fixed, the receptive field is also fixed. But different household appliances may have different operating states and runtimes, thus requiring different receptive fields. Summary of the Invention

[0004] The purpose of this invention is to provide a non-intrusive load decomposition model and method for spatiotemporal feature-based attention fusion, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A non-intrusive load decomposition model based on spatiotemporal feature fusion attention is proposed. The model extracts the temporal features of the load input sequence through a temporal feature extraction module and the spatial features of the load input sequence through a spatial feature extraction module. The temporal and spatial features are concatenated and input into a convolutional attention module for weighted learning and fusion of spatiotemporal features. Finally, the fused features are used to calculate the power consumption of each load through a fully connected layer.

[0007] Preferably, the load input sequence is the total power consumption of the house based on a time series.

[0008] Preferably, the time feature extraction module is a bidirectional long short-term memory network.

[0009] Preferably, the spatial feature extraction module is an Inception module, which extracts multi-scale spatial features of the load input sequence.

[0010] Preferably, the inception module includes a plurality of parallel convolutional units, each of which includes at least one cascaded one-dimensional convolution.

[0011] Preferably, the convolutional attention module extracts joint channel and spatial attention features based on cascaded temporal and spatial features, and multiplies these attention features with the load input sequence.

[0012] A non-intrusive load decomposition method for spatiotemporal feature-based attention fusion, the method comprising:

[0013] The non-intrusive load decomposition model for spatiotemporal feature-based attention fusion is trained.

[0014] The power consumption of each load is obtained by using a non-intrusive load decomposition model that integrates attention based on spatiotemporal features after training.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1) This invention combines temporal and spatial features and performs spatiotemporal feature fusion learning through a convolutional attention module. The convolutional attention module combines temporal attention and channel attention to perform spatiotemporal feature fusion learning, which facilitates effective feature extraction and reduces the impact or coverage of features by high-power home appliances.

[0017] 2) This invention uses an inception module constructed by multiple parallel convolutions of different sizes. This inception module is used as a spatial feature extraction module to extract features from different receptive fields, thereby improving the accuracy of prediction. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a non-intrusive load decomposition model structure for spatiotemporal feature fusion attention according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the BiLSTM structure of the present invention.

[0020] Figure 3 This is a flowchart of the convolutional attention module.

[0021] Figure 4 The results are from ablation experiments in a visible scenario.

[0022] Figure 5This presents a comparison of different load decomposition methods in the visible scenario.

[0023] Figure 6 This is a comparison of different load decomposition methods in invisible scenarios. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0025] A non-intrusive load decomposition model based on spatiotemporal feature fusion attention is proposed. The model extracts the temporal features of the load input sequence through a temporal feature extraction module and the spatial features of the load input sequence through a spatial feature extraction module. The temporal and spatial features are concatenated and input into a convolutional attention module for weighted learning and fusion of spatiotemporal features. Finally, the fused features are used to calculate the power consumption of each load through a fully connected layer.

[0026] The load input sequence is the total power consumption of the house based on a time series.

[0027] The time feature extraction module is a bidirectional long short-term memory network.

[0028] In this invention, since household appliances follow repetitive patterns over long periods, short-term dependencies are insufficient. For example, a washing machine follows several operating cycles (such as pre-wash, washing, drying, etc.), each cycle being interconnected over a relatively long period. Therefore, temporal feature extraction is implemented using a bidirectional long short-term memory (BiLSTM) network. The structure of BiLSTM is as follows... Figure 2 As shown, from Figure 2 As can be seen from this, its output is determined by the forward and backward states, that is... Through the forward channel, neurons can extract features relevant to the previous time step; through the backward channel, neurons can extract features relevant to the next time step. This makes the temporal features extracted by BiLSTM richer, which is very beneficial for improving the accuracy of load decomposition.

[0029] The spatial feature extraction module is the Inception module, which extracts multi-scale spatial features of the load input sequence.

[0030] The Inception module includes several parallel convolutional units, each of which includes at least one cascaded one-dimensional convolution.

[0031] In this invention, since different household appliances may have different operating states and operating times, selecting only one convolution operation with a fixed receptive field would not be sufficient to consider the characteristics of all appliances. For example, loads with short operating times require a small receiving electric field, while loads with long operating times require a large receiving electric field. Therefore, this invention employs multiple parallel convolutions to extract features from different receptive fields. Because the load input sequence is one-dimensional, the inception module used is composed of one-dimensional convolutions, and the parallel convolution units also preserve the channel characteristics of the entire inception block.

[0032] In this embodiment, refer to Figure 1 As shown, the Inception module consists of five parallel convolutional neural networks of different sizes, each serving as a convolutional unit. The first convolutional neural network (convolutional unit) includes sequentially connected max-pooling layers and convolutional layers. The max-pooling layer has a 3*1 kernel size, and the convolutional layer has a 64*1 kernel size. The structures of the remaining four convolutional neural networks are directly referenced. Figure 1 This will not be elaborated upon here.

[0033] The convolutional attention module extracts joint channel and spatial attention features based on cascaded temporal and spatial features, and multiplies these attention features with the load input sequence.

[0034] In this invention, the convolutional block attention module (CBAM) is an attention network composed of channel attention and spatial attention. Figure 3 The flowchart of the convolutional attention module is shown, see reference. Figure 3As shown, for cascaded spatial and temporal features, global average pooling and global max pooling features are first obtained through global average pooling and global max pooling branches, respectively. These two features are then subjected to two convolution operations: the first convolution compresses the channels by f / 4, and the second convolution expands the channels by f. These two features are then concatenated and passed through a convolutional layer and a ReLU activation function to obtain an output. This output is multiplied by the cascaded spatial and temporal features to obtain a channel attention map. This channel attention map is then passed through an average pooling layer and a max pooling layer to obtain average pooling and max pooling features, respectively. These average pooling and max pooling features are concatenated and then passed through a convolutional layer and a sigmoid activation function to obtain a spatial attention map. This spatial attention map is multiplied by the channel attention map to obtain a joint channel and spatial attention map. In convolutional neural networks, channel information generally represents different feature information of an image; therefore, by selecting channels, the network can better notice information in the image that is useful for the task.

[0035] In this invention, for a given intermediate feature map f, the CBAM module sequentially learns channel and spatial attention maps along two independent dimensions (channel and spatial), and then multiplies the resulting spatial attention map with the channel attention map for adaptive feature optimization. This operation emphasizes the importance of different features from two different dimensions. Furthermore, CBAM is a lightweight and general-purpose module, so its computational cost is negligible. Moreover, it can be freely inserted into any CNN architecture and can be trained end-to-end with the base CNN.

[0036] The calculation expression for CBAM is as follows:

[0037]

[0038] f cs =f c *σ(conv1d(Avgpool(f c )+Maxpool(f c (2)

[0039] Among them, f c and f cs These represent the channel attention map and the joint channel and spatial attention map, respectively. σ represents the sigmoid activation function.

[0040] Since the input signal is one-dimensional, one-dimensional convolution is used to extract features. Specifically, the first-layer convolutional filter size in the channel attention module is f (the number of channels) divided by the reduction ratio. In this embodiment, the reduction ratio is set to 4. The filter sizes in the second-layer convolution and the spatial attention convolutions are both f (the number of channels), and all convolutional kernel sizes and strides in the attention mechanism are 1.

[0041] In this embodiment, a total power sliding window is used to extract different types of features. First, a spatial feature extraction module composed of multiple parallel convolutions (with different kernel sizes) is used to extract spatial features at different scales. Let f be the feature extracted by each convolutional unit. i Let i = 1, 2, 3, 4, 5, where i represents the number of parallel convolutional units. Then the output of the spatial feature extraction module is:

[0042] F t =[f1, f2, f3, f4, f5] (3)

[0043] Simultaneously, BiLSTM is used to process the time information in the input sequence, and finally the output result r of BiLSTM is derived. t Then, the input spatial and temporal features are concatenated using f. t =[F t r t The cascaded features contain rich spatiotemporal characteristics, which are then fed into a convolutional attention module consisting of channel and spatial attention for weighted fusion learning to eliminate the influence of feature occlusion. Finally, the weighted fused features are input into a fully connected layer to calculate the prediction point for each load.

[0044] Based on this, the present invention discloses a non-intrusive load decomposition method for spatiotemporal feature-based attention fusion, the method comprising:

[0045] The non-intrusive load decomposition model for spatiotemporal feature-based attention fusion is trained.

[0046] The power consumption of each load is obtained by using a non-intrusive load decomposition model that integrates attention based on spatiotemporal features after training.

[0047] In this embodiment, the model is trained based on a dataset, which includes several samples, each of which includes an input sequence and the corresponding power consumption of each load.

[0048] To evaluate the effectiveness of the proposed method, experimental simulation tests were conducted using the publicly available dataset UKDALE.

[0049] This dataset contains monitoring data on total electricity consumption and household electricity usage recorded every 6 seconds for five UK homes. The appliances used in the training and testing experiments included a kettle, microwave, refrigerator, dishwasher, and washing machine. This invention conducted two experimental tests on these appliances: one in visible scenarios and one in invisible scenarios. For the visible scenarios, 80% of the data extracted from room 1 was used as the training set, and the remaining 20% ​​as the test set. For the invisible scenarios, room 1 was used as the training set, and room 2 as the test set.

[0050] Since the data in the UKDALE dataset were collected from real-world scenarios, issues such as inconsistent sampling and missing data may arise. To obtain a complete and uniform dataset, all data in the UKDALE dataset were downsampled to 8 seconds, ensuring that the sampling time for each appliance corresponds one-to-one with its total power. Furthermore, to eliminate the influence of different units or scales, the data was converted to dimensionless values, and the input data underwent normalization preprocessing.

[0051] The batch size during training was 128, the dropout value was 0.5, and the loss function was mean squared error. The Adam optimizer was used for training with a learning rate of 0.001. The maximum number of epochs was 100, and the patience was 10.

[0052] Two typical evaluation metrics, mean absolute error (MAE) and signal aggregation error (SAE), are used to evaluate the performance of different methods. MAE measures the average power error at each time point, as shown below:

[0053]

[0054] In the formula, Let p be the predicted value of the i-th load at time t. i (t) represents its corresponding ground truth value. T is the number of time points. SAE is the total energy error (SAE) used to measure the total energy consumption error over a period of time, as shown below:

[0055]

[0056] In the formula, p i Let i be the total power consumption of the i-th load. For the corresponding p i The predicted estimate.

[0057] This invention conducted three different ablation experiments in both visible and invisible scenarios: the Inception spatial feature extraction module, the BiLSTM temporal feature extraction module, and the STFF method without an attention module. Inception refers to considering the impact of spatial feature extraction alone, BiLSTM refers to considering the impact of temporal feature extraction alone, and STFF represents a spatiotemporal feature fusion method without convolutional attention blocks (CBAM), simply concatenating spatiotemporal features for subsequent load decomposition.

[0058] For the ablation experiment results in the visible scene, as follows: Figure 4 As shown. By Figure 4 As can be seen, the method of this invention achieves the minimum MAE and SAE, and its three modules each play a different role in load decomposition. Specifically, BiLSTM outperforms Inception, indicating that the temporal feature extraction module BiLSTM is slightly more effective than the spatial feature extraction module Inception. In particular, the experimental results for the kettle and microwave oven loads show that BiLSTM performs better. This suggests that temporal features are more important for these two loads. Furthermore, the method of this invention significantly outperforms the STFF method without a convolutional attention network. This demonstrates that the convolutional attention block (CBAM) has a positive impact on spatiotemporal feature fusion, thus improving the performance of load decomposition.

[0059] Figure 5 The comparison results of different load decomposition methods in the visible scenario are shown, by Figure 5 The superiority of the method of this invention is evident. The main reason is that the method of this invention focuses on attention learning of spatiotemporal features, while the Seq2point and BitcnNILM methods only focus on learning a single temporal or spatial feature.

[0060] Figure 6 A comparison of different load decomposition methods in invisible scenarios is presented. Figure 6 As can be seen, the MAE and SAE scores obtained by the method of this invention are superior to those of the Seq2point and BitcnNILM methods. This indicates that the method of this invention is also effective for load decomposition in invisible scenarios.

Claims

1. A non-intrusive load decomposition model for spatiotemporal feature-based attention fusion, characterized in that, The temporal features of the load input sequence are extracted by the temporal feature extraction module, and the spatial features of the load input sequence are extracted by the spatial feature extraction module. The temporal and spatial features are concatenated and input into the convolutional attention module for weighted learning and fusion of the spatiotemporal features. Finally, the fused features are used to calculate the power consumption of each load through a fully connected layer. The load input sequence is the total power consumption of the house based on a time series. The time feature extraction module is a bidirectional long short-term memory network. The spatial feature extraction module is an inception module, which extracts multi-scale spatial features of the load input sequence. The inception module includes several parallel convolutional units, each of which includes at least one cascaded one-dimensional convolution. The convolutional attention module extracts joint channel and spatial attention features based on cascaded time and spatial features, and multiplies these attention features with the load input sequence. First, a spatial feature extraction module composed of multiple parallel convolutions is used to extract spatial features at different scales. Let f be the feature extracted by each convolutional unit. i Let i = 1, 2, 3, 4, 5, where i represents the number of parallel convolutional units. Then the output of the spatial feature extraction module is: F t =[f1, f2, f3, f4, f5], Simultaneously, BiLSTM is used to process the time information in the input sequence, and finally the output result r of BiLSTM is derived. t Then, the input spatial and temporal features are concatenated using f. t =[F t r t ]; For the cascaded spatial and temporal features, firstly, global average pooling and global max pooling features are obtained through global average pooling and global max pooling branches, respectively. These two features are then subjected to two convolution operations: the first convolution compresses the channels by f / 4, and the second convolution expands the channels by f. These two features are then concatenated and passed sequentially through a convolutional layer and a ReLU activation function to obtain an output. This output is multiplied by the cascaded spatial and temporal features to obtain a channel attention map. This channel attention map is then passed through an average pooling layer and a max pooling layer to obtain average pooling and max pooling features, respectively. These average pooling and max pooling features are concatenated, and then passed through a convolutional layer and a sigmoid activation function to obtain a spatial attention map. This spatial attention map is multiplied by the channel attention map to obtain a joint channel and spatial attention map. The calculation expression for the Convolutional Attention Module (CBAM) is shown below: f c =f*(Relu(convld(convld(Avgpool(f)))+convld(convld(Avgpool(f))))) (1), f cs =f c *σ(conv1d(Avgpool(f c )+Maxpool(f c ))) (2), Among them, f c and f cs Let represent the channel attention map and the joint channel and spatial attention map, respectively, and σ represent the sigmoid activation function.

2. A non-intrusive load decomposition method for attention fusion based on spatiotemporal features, characterized in that, The method includes: The non-intrusive load decomposition model for spatiotemporal feature fusion attention as described in claim 1 is trained; The power consumption of each load is obtained by using a non-intrusive load decomposition model that integrates attention based on spatiotemporal features after training.

Images