Intelligent sensing method and device for enhancing environment self-adaptive application electric behavior based on multi-task learning

Abstract

The application discloses an environment self-adaptive power consumption behavior enhancement intelligent sensing method and device based on multi-task learning, collects load bus current data, performs short-time Fourier transform on the current data, takes time-frequency information under all frequency bands of the obtained time-frequency distribution as extracted features, inputs the features into a multi-task learning improved time convolution network, so as to simultaneously obtain output results of multiple tasks of power consumption behavior category identification, power consumption behavior type subdivision and power consumption behavior prediction, controls corresponding load switching according to the power consumption behavior prediction result, and compares and verifies the newly obtained power consumption behavior type subdivision result with the previous power consumption behavior prediction result. The application can improve intelligent power consumption experience in a multi-load power consumption scene and avoid safety hazards caused by equipment operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-invasive multi-load multi-switching power consumption behavior recognition and detection, and involves using a deep learning neural network to synchronously complete multi-scale classification of power consumption behavior categories and types and power consumption behavior prediction. Specifically, it involves applying time-frequency transformation processing to the same current signal, screening features, and importing them into a multi-task learning improved temporal convolutional network for training to accurately identify the load type, perform classification at a coarser scale, and conduct comparative verification of power consumption behavior prediction. Background Art

[0002] As the proportion of electricity in terminal energy consumption has been increasing year by year, the demand scale for collecting power consumption behavior information has been continuously expanding. At the same time, relevant emerging electricity technologies are also extending to the distribution network operating under higher voltage and current levels, and the combined application scale in the power distribution and consumption links accounts for more than 50%. Non-invasive load monitoring technology can monitor information such as the type, operating conditions, and related parameters of each household load through the analysis of the electricity quantity at the power supply entrance of the power bus, thereby significantly reducing the cost of smart electricity infrastructure and providing effective support for the power grid to revise the power distribution and consumption terminal regulation strategy.

[0003] There are a rich variety of electrical products in the smart home market, such as smart lights, security systems, smart bathrooms, etc. To enhance the interactive connection and coordinated control of various home devices, non-invasive load monitoring technology has entered the hot research field. Non-invasive load monitoring technology has many positive effects on energy conservation and emission reduction of smart home products. Through the feedback of users' electricity consumption details, the maximum power saving potential can reach 35%. For example, the smart home lighting system can make full use of the lighting effect of natural light, and while ensuring the indoor brightness, reduce the energy consumption to the lowest level by automatically adjusting the brightness of the lights. However, with the significant increase in the demand for multi-load联动 intelligent power consumption experience in the home environment, only providing classification of power consumption behavior at a finer scale through non-invasive load monitoring technology will cause certain limitations in smart home services in some specific scenarios. At the same time, existing power consumption behavior prediction models are mainly used for classifying users' current power consumption behavior, such as Chinese patents CN115641228A and CN113901977A.

[0004] Customized electricity usage behavior adjustment and management can not only adjust electricity usage habits to minimize energy consumption and reduce electricity costs, but also meet users' electricity management needs and reduce the risk of electric shock due to problems such as aging insulation or loose connections when contacting loads (i.e., electrical products). For example, it can automatically open and close the garage, turn on the air conditioner, water heater, and audio system, and switch lights when the user returns home at night, or provide users with specific time-limited services such as cooking (e.g., turning on the rice cooker), stir-frying (e.g., turning on the microwave), washing dishes (e.g., turning on the dishwasher), doing laundry (e.g., turning on the washing machine), and opening and closing windows. However, these customized services are usually provided to users based on environmental scenarios and demographic characteristics, analyzing their living habits and providing them to users within specific time periods. Therefore, it is difficult to efficiently and in real-time update them to meet the changing needs of users' electricity management.

[0005] Furthermore, compared to the load type classification scale involved in electricity consumption behavior classification, the classification scales involved in classifying electricity theft on dedicated lines (such as the grid connection status of non-coal-to-electricity appliances), illegal use of electrical appliances (such as the operation status of resistive heating loads on campuses), and the shutdown of industrial electrical equipment under time-limited power rationing during unplanned production or abnormal working conditions are relatively coarse. Faced with the objective need for multi-scale classification of electricity consumption behavior, it is necessary to address the difficulty in using a unified algorithm framework to construct classification models due to differences in the number of classification states, the number of features, and the complexity of model computation. At the same time, these differences also make it difficult for multi-scale electricity consumption behavior classification methods to share measurement signals and computational frameworks with existing technologies.

[0006] In summary, there is an urgent need to develop an environment-adaptive (customer-dependent) enhanced perception algorithm framework for electricity behavior that integrates multiple tasks such as electricity behavior category identification, type segmentation, and electricity behavior prediction, making it possible to improve the computational efficiency of the algorithm by applying multi-task learning. Summary of the Invention

[0007] The purpose of this invention is to provide an environment-adaptive power consumption behavior enhancement intelligent perception method and device based on multi-task learning. It can accurately and quickly classify power consumption behavior at multiple scales and predict power consumption behavior, thereby providing real-time and intelligent power consumption solutions that meet user needs.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An environmental adaptive power consumption behavior enhancement intelligent sensing method includes the following steps:

[0010] 1) The current signal output from the load bus is measured at a sampling frequency using a current sensor. fs Perform point-by-point sampling to obtain the first... jLoad bus current data for each analysis period x j Proceed to step 2 for load bus current data characteristic analysis;

[0011] 2) Collect the load bus current data x j Perform time-frequency transformation to obtain the time-frequency distribution of the current data, extract the time-frequency information of each frequency band in the time-frequency distribution to obtain the input features, and proceed to step 3.

[0012] 3) The multi-task learning model (i.e., the multi-task learning improved deep learning neural network) trained by the fusion deep learning neural network performs load type identification and classification on the input features at multiple scales, thereby simultaneously obtaining the output results of electricity behavior category identification and electricity behavior type subdivision as well as the output results of electricity behavior prediction at the corresponding scale, and then proceeds to step 4.

[0013] 4) According to the first j The electricity consumption prediction results obtained from the first analysis period will be used to remotely control the corresponding load in the next analysis period (the second analysis period). j The switching was completed within the +1 analysis period. Based on the electricity behavior type segmentation results and electricity behavior category identification results obtained in this analysis period, the previous analysis period (the +1 analysis period) was re-processed. j The predicted electricity consumption behavior results obtained from the analysis period were compared and verified.

[0014] Preferably, the sampling frequency fs The frequency range is 25~60kHz, and the number of sampling points during the analysis period is ≥3000.

[0015] Preferably, the time-frequency transformation adopts a short-time Fourier transform with a rectangular window, wherein the window length is selected from 512 to 2048 based on a trade-off between the speed and effectiveness of information processing.

[0016] Preferably, in the multi-task learning model that integrates deep learning neural networks, the deep learning neural network adopts a temporal convolutional network.

[0017] Preferably, the multi-task learning model that integrates deep learning neural networks is achieved by modifying the fully connected layers in the temporal convolutional network, thereby combining a multi-task learning model based on the Keras-Shared-Bottom model into the temporal convolutional network. The output of the multi-task learning model that integrates deep learning neural networks is represented as follows:

[0018]

[0019] In the formula, h t The input at time t, yk For the first k The output obtained from each task k =1,2…, n , n For the number of tasks, b n This represents the deviation vector for each task. w n This is the weight matrix for each task.

[0020] Preferably, the training of the multi-task learning model fused with deep learning neural networks includes the following steps: collecting current data of the load to be identified during operation according to the length of the analysis period, performing corresponding time-frequency transformation, filtering the frequency bands of the obtained time-frequency distribution, and selecting frequency bands according to the requirements that the time-frequency information of the selected frequency bands under each classification scale (for electricity behavior category identification or electricity behavior type subdivision, each electricity behavior classification state is determined by the corresponding electricity behavior classification scale) meets the requirements that multiple electricity behaviors under the same classification state can obtain common electricity behavior feature parameters and / or multiple electricity behaviors under different classification states can obtain significantly different electricity behavior feature parameters. The time-frequency information under this frequency band is used as the training feature for electricity behavior category identification or electricity behavior type subdivision, and imported into the multi-task learning model fused with deep learning neural networks.

[0021] Preferably, the screening specifically includes the following steps: performing short-time Fourier transform processing on the current data of the load to be identified during operation, and then randomly selecting time-frequency information from one or more frequency bands from the obtained time-frequency distribution.

[0022] Preferably, the feature parameter is one or more of the mean, variance, and variance variation of time-frequency information under the selected frequency band.

[0023] Preferably, in step 4, after the corresponding load is switched on and off in the next analysis period, if the electricity behavior type subdivision result or electricity behavior category identification result obtained in the analysis period is inconsistent with the electricity behavior prediction result obtained in the previous analysis period, that is, the electricity behavior prediction result is not verified correctly, then the electricity behavior prediction result obtained in the corresponding analysis period is changed and iterative training is performed until the electricity behavior prediction result is consistent with the electricity behavior type subdivision result and electricity behavior category identification result obtained in the prediction analysis period, that is, the electricity behavior prediction result is verified correctly, then the iterative training is stopped.

[0024] Preferably, the number of iterations for training is 100 to 200.

[0025] An environmentally adaptive power consumption behavior enhancement intelligent sensing device includes a detection system based on the above-mentioned environmentally adaptive power consumption behavior enhancement intelligent sensing method. The detection system includes a load bus output current signal sampling module, a load bus current data feature analysis module, a multi-scale power consumption behavior classification and power consumption behavior prediction module, and a comparison and verification module.

[0026] The load bus output current signal sampling module is used to sample the load bus output current signal at a sampling frequency using a current sensor. fs Perform point-by-point sampling until the load bus current data for one analysis period is obtained;

[0027] The load bus current data feature analysis module is used to analyze the collected data. j Load bus current data for each analysis period x j Perform time-frequency transformation and extract the time-frequency information of each frequency band in the time-frequency distribution obtained by time-frequency transformation as input features;

[0028] The multi-scale electricity consumption behavior classification and electricity consumption behavior prediction module is used to use a multi-task learning model (i.e., a multi-task learning improved deep learning neural network) trained by the fusion deep learning neural network to identify and classify load types at multiple scales and predict electricity consumption behavior at the corresponding scales, thereby simultaneously obtaining the output results of electricity consumption behavior category identification, electricity consumption behavior type subdivision and electricity consumption behavior prediction.

[0029] The comparison and verification module is used to compare and verify based on the first j The electricity consumption prediction results obtained from the first analysis period will be used to remotely control the corresponding load in the next analysis period (the second analysis period). j The switching was completed within the +1 analysis period, and the previous analysis period (the +1 analysis period) was further refined based on the detailed breakdown of electricity consumption behavior types and the identification results of electricity consumption behavior categories obtained during this analysis period. j The predicted electricity consumption behavior results obtained from the analysis period were compared and verified.

[0030] Preferably, the detection system further includes a model training module. This module is used to collect current data of the load to be identified during operation according to the analysis period length, perform corresponding time-frequency transformation, filter the frequency bands of the obtained time-frequency distribution, and select frequency bands according to the requirements that the time-frequency information of the selected frequency bands under each classification scale (for electricity behavior category identification or electricity behavior type subdivision, the number of electricity behavior classification states is divided according to the corresponding electricity behavior classification scale) meets the requirements that multiple electricity behaviors under the same classification state can obtain common electricity behavior feature parameters and / or multiple electricity behaviors under different classification states can obtain significantly different electricity behavior feature parameters. The module also uses the time-frequency information under the frequency band as the training feature for electricity behavior category identification or electricity behavior type subdivision, and imports it into a multi-task learning model that integrates deep learning neural networks for training.

[0031] The beneficial effects of this invention are reflected in:

[0032] This invention performs time-frequency transformation on the same current signal (load bus output current signal) and uses the time-frequency information of the current signal as input features. It then utilizes a deep learning neural network improved with a multi-task learning model to identify and subdivide the electricity consumption behavior categories of different loads. Based on accurate identification of different load types and coarser-scale classification, it compares and verifies the newly obtained electricity consumption behavior subdivision results with previous electricity consumption behavior prediction results. This unified algorithm framework enables electricity consumption behavior perception at multiple classification scales and bidirectional interaction with the environment, significantly improving the adaptability of electricity consumption behavior perception to the environment and making intelligent load control possible. This invention not only quickly identifies loads but also provides optimal load operation schemes based on the user's own electricity consumption habits, reducing power consumption and improving the intelligent electricity consumption experience in multi-load scenarios, while avoiding safety hazards in load equipment operation and improving electricity safety.

[0033] Furthermore, the electricity consumption behavior category identification and electricity consumption behavior type subdivision in this invention are components of the multi-scale (large and small) classification characterization of electricity consumption behavior. Therefore, both the input features and training features are obtained by processing the current signal through short-time Fourier transform. This makes it easier to select the frequency band that meets the principle of maximizing the significance of the multi-scale classification feature parameters of electricity consumption behavior during training, and to obtain better training results.

[0034] Furthermore, when training the deep learning neural network (specifically a temporal convolutional network) improved by the multi-task learning model, this invention extracts and imports different feature sets according to the requirements of different scales of electricity consumption behavior classification tasks. This enables the multi-scale classification model obtained through joint training to have the ability to predict future electricity consumption behavior, and maintains an accuracy of over 95% for both multi-scale classification of current electricity consumption behavior and prediction of future electricity consumption behavior.

[0035] Furthermore, the multi-task learning model that integrates deep learning neural networks in this invention adopts the Keras-Shared-Bottom model framework, which can combine the electricity consumption behavior prediction results to obtain the electricity consumption behavior type subdivision prediction results, thereby correcting and improving the accuracy of the type subdivision results. Attached Figure Description

[0036] Figure 1 This is a flowchart of the environmental adaptive power consumption behavior enhancement intelligent sensing method in an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the implementation principle of a bidirectional adaptive system.

[0038] Figure 3a This is one of the exemplary current waveforms of five electrical appliances collected under the same operating time (20s).

[0039] Figure 3b The current feature map is used to identify the electricity consumption behavior categories in the selected frequency band (8.4kHz) after the current data undergoes short-time Fourier transform.

[0040] Figure 3c To combine the electricity consumption behavior category identification map with multi-task learning (test results of electricity consumption behavior category identification training alone: ​​the output is low level 0 when no load is applied, and the output is different high level 1 and 2 when the corresponding type of load is applied).

[0041] Figure 4a This is the second example of the current waveforms of five electrical appliances collected under the same operating time (20s).

[0042] Figure 4b This is a subdivided current characteristic map of electricity consumption behavior types in the selected frequency band (194Hz) after the current data undergoes short-time Fourier transform.

[0043] Figure 4c To combine multi-task learning to create a segmentation map of electricity consumption behavior types (test results of training on individual electricity consumption behavior type segmentation: the output is low level 0 when no load is applied, and different high levels 1, 2, 3, 4, 5 when the corresponding load is applied).

[0044] Figure 5a The prediction graph for electricity behavior category identification combined with multi-task learning (test results of electricity behavior category identification prediction after joint training of electricity behavior category identification and electricity behavior type subdivision).

[0045] Figure 5bThis is a prediction map for electricity behavior type segmentation that combines multi-task learning (test results of electricity behavior type segmentation prediction after joint training of electricity behavior category identification and electricity behavior type segmentation). Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments described are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0047] Combination Figure 1 This invention elucidates the steps of a multi-task learning-based method for enhancing intelligent perception of environmental adaptive electricity consumption behavior, which utilizes short-time Fourier transform and is trained using a temporal convolutional network (TCN). Specific details are as follows:

[0048] Step 1: Output current signal to the load bus using a current sensor at a sampling frequency. fs ( fs The load bus current data is obtained by sampling point by point at 50kHz. x j ,in j For analyzing time period numbers, j ∈N+, when x j Once the number of sampling points reaches the requirement for the analysis period (≥3000 sampling points), proceed to step two for load bus current data characteristic analysis.

[0049] Step 2: Perform a short-time Fourier transform on the load bus current data collected during the current analysis period using the following formula (the choice of time-frequency transformation method is a trade-off between comprehensive significance and hardware computing power) to obtain the time-frequency distribution of the current data:

[0050]

[0051] Where j represents the complex unit, and s(n) is the input current signal (specifically, the load bus current data mentioned above). N For window length (e.g.) N (Take 1024), where n is the window function. w A rectangular window containing only 1s. f For frequency ( f Values ​​and fs Same, i.e., 50kHz). k The center point of the window function.

[0052] The time-frequency distribution of the obtained current data is plotted using MATLAB to generate the corresponding original feature map. That is, all time-frequency information under each frequency band of the time-frequency distribution obtained by time-frequency transformation is used as input features, and then proceed to step three.

[0053] Step 3 involves inputting the extracted features (i.e., input features) into a multi-task learning model of a fusion temporal convolutional network that has been trained. This model then simultaneously outputs the results of electricity consumption behavior category identification and electricity consumption behavior type subdivision for the current analysis period, and proceeds to Step 4.

[0054] The aforementioned multi-task learning model incorporating a temporal convolutional network combines a multi-task learning model based on the Keras-Shared-Bottom model with a temporal convolutional network. By modifying the fully connected layers in the temporal convolutional network, this improved deep learning neural network can perform multi-task outputs. The formula for the original temporal convolutional network is:

[0055]

[0056] In the formula, To expand the causal operator, d To expand speed (control jump distance). For related convolution kernels, superscript l This represents the number of layers in the convolutional network. and The subscripts indicate two different layers (distinguished as 1 and 0) in the residual block structure of the temporal convolutional network. y l For input, y l+1 For output, y l+1 That is, the following " t In the output of time" h t .

[0057] The formula for the ReLU activation function is:

[0058]

[0059] In the formula, w The output of the corresponding layer is used as the input to the activation function.

[0060] The above temporal convolutional network t The output expression at time t is:

[0061]

[0062] In the formula, y t For the output, the fully connected layers in the temporal convolutional network use the Softmax function as the activation function, and b is the bias vector. h t The input at time t ( h tThe above is the initial input to the model. y l (This is the input to a fully connected layer for multiple tasks).

[0063] The Softmax function is explained as follows:

[0064]

[0065] In the formula, z i For the first i The output values ​​of each of the C nodes are C, where C is the number of output nodes, i.e., the number of categories. The Softmax function can be used to convert the C output values ​​into a probability distribution where each state value is in the range [0,1] and the sum of the corresponding state values ​​for all output values ​​is 1.

[0066] The aforementioned "changing the fully connected layers of the temporal convolutional network" is the core of the Keras-Shared-Bottom model framework, which integrates temporal convolutional networks to synchronously achieve multi-task output. This allows the temporal convolutional network to go beyond simply outputting the result of a single task at time t (outputting y). t After improvement, it can achieve simultaneous output of the results of multiple tasks (output is y). k The output layer formula is changed to:

[0067]

[0068] In the formula, h t The input at time t, y k For the first k The output obtained from each task k =1,2…, n , n The number of state recognition tasks (including electricity consumption behavior prediction tasks) to be performed. b n This represents the deviation vector for each task. w n This is the weight matrix for each task.

[0069] The construction principle of the multi-task learning is to use a single expert system framework with multiple inputs and multiple outputs that minimizes model storage and computation. The parameters of each task in the framework have the ability to be shared. The shared parameters of electricity behavior category identification are used to obtain the equipment operating environment and type attributes as further subdivision constraints. Combined with the electricity behavior prediction results, the subdivision prediction results of electricity behavior types are obtained, thereby correcting and improving the accuracy of the subdivision results of electricity behavior types.

[0070] Since electricity consumption behavior category identification and electricity consumption behavior type segmentation are components of multi-scale (large and small) classification of electricity consumption behavior, the time-frequency distribution of the obtained current data is processed by frequency band filtering according to electricity consumption behavior category identification (e.g., classifying the identified different types of loads into resistive heating loads and non-resistive heating loads) and electricity consumption behavior type segmentation (i.e., identifying different types of loads), thereby extracting their respective training features. The training features are then combined with the load state labels of the corresponding classification scale to form a training set, which is then input into the improved temporal convolutional network (i.e., a multi-task learning model that integrates temporal convolutional networks) for training.

[0071] After frequency band filtering, the number of electricity consumption behavior categories is divided according to the required electricity consumption behavior classification scale. The characteristic parameters (time-frequency information of loads belonging to this category in a specific frequency band obtained through frequency band filtering) of multiple electricity consumption behaviors within the same category have commonalities. The criterion for "commonalities" is: within the same analysis period (i.e., at least 3000 sampling points), the characteristic error of the electricity consumption behavior characteristic parameters (mean or variance of time-frequency information in the selected frequency band) is no greater than 5% (the characteristic error is calculated by subtracting each pair of characteristic parameters and dividing by the larger of the two). The characteristic parameters of multiple electricity consumption behaviors within different category states have significant differences. The criterion for "significant differences" is: within the same analysis period (i.e., at least 3000 sampling points), the characteristic error of the electricity consumption behavior characteristic parameters (changes in the mean or variance of time-frequency information in the selected frequency band) is greater than 30%. Different electricity consumption behaviors will have some "significantly different" electricity consumption behavior feature parameters and some "common" electricity consumption behavior feature parameters. The part of the electricity consumption behavior feature parameters that are extracted as feature parameters for electricity consumption behavior category identification can be used to complete the major category identification, while the part of the feature parameters that are "significantly different" can be used as feature parameters for electricity consumption behavior type subdivision to complete the minor category identification.

[0072] Step four involves inputting the extracted features (i.e., input features) into a multi-task learning model of a fusion temporal convolutional network that has been trained. This model can then output the predicted power consumption behavior for the next analysis period based on the learned user power consumption behavior (e.g., load status identification results of each day's power consumption behavior type in a quarterly period). The predicted power consumption behavior can be used to remotely control the switching of the corresponding loads (appliances) before the start of the next analysis period.

[0073] To learn user electricity consumption behavior, after entering the next analysis period, the scale identification results of electricity consumption behavior corresponding to the analysis period obtained in step three (such as the subdivision results of new electricity consumption behavior types) are used to compare and verify the electricity consumption behavior prediction results output by the model in the previous analysis period. If the predicted electricity consumption behavior is inconsistent with the user's actual electricity consumption behavior (i.e., the verification is incorrect), the electricity consumption behavior control switching signal is changed (that is, the prediction results output in the previous analysis period are modified, so as to carry out iterative training of the temporal convolutional network) until the electricity consumption behavior prediction results are verified to be correct.

[0074] The aforementioned environmentally adaptive electricity consumption behavior enhancement intelligent perception method based on multi-task learning employs a multi-task learning model with the Keras-Shared-Bottom model as its main framework to improve upon the shortcomings of TCN's single-task execution. This allows for simultaneous identification of electricity consumption behavior categories and segmentation of electricity consumption behavior types based on extracted features. Iterative training is then performed on the basis of segmented electricity consumption behavior types to achieve accurate electricity consumption behavior prediction, further accelerating the speed of electricity consumption behavior category identification, segmentation, and prediction. Multiple tasks can be completed in a single analysis, reducing time, computation, storage, and even model maintenance costs. Furthermore, by addressing the shortcomings of TCN's single-task execution, multiple tasks with relatively sparse datasets are learned within a single model, mitigating overfitting and improving the model's generalization ability to some extent.

[0075] like Figure 2As shown, existing electricity behavior sensing technologies mainly extract users' electricity behavior information to classify electricity behavior types. This results in smart home devices failing to significantly improve user well-being through autonomous control, still requiring manual load switching (e.g., load switching in multi-load linked smart electricity application scenarios). In other words, existing smart homes are not truly "smart." The proposed invention, an environment-adaptive enhanced intelligent sensing method based on multi-task learning, integrates with existing electricity behavior sensing devices, reducing memory usage and enabling adaptive enhanced intelligent sensing (forming an adaptive enhanced intelligent sensing device). Specifically, it uses extracted behavior information from various loads to classify electricity behavior types, stores operational information and patterns of electricity behavior for learning, and simultaneously identifies the category of electricity behavior according to a larger classification scale, completing the task of identifying the major category to which the load belongs, and remotely controlling future load switching based on prediction results. Furthermore, by analyzing and comparing the information obtained after remote control of the electrical load based on the user's actual electricity consumption behavior (since the prediction results relied upon by remote control may be incorrect, the user will naturally switch on the corresponding electrical load to correct the error upon discovering the incorrect switching), the resulting breakdown of electricity consumption behavior types and identification results are compared and verified with the previous electricity consumption behavior prediction results. If the verification result shows that the user's actual electricity consumption behavior is consistent with the predicted electricity consumption behavior, subsequent adaptive electricity consumption behavior enhancement intelligent sensing continues. If the verification result shows a deviation, the previous erroneous prediction results are corrected and recorded, and then the behavioral information of the electrical load is further learned to improve the accuracy of electricity consumption behavior prediction. This forms a complete closed-loop system that combines electricity consumption behavior type breakdown, electricity consumption behavior category identification, electricity consumption behavior prediction, and correction. After long-term learning of the usage patterns and behavioral information of the electrical load, the adaptive electricity consumption behavior enhancement intelligent sensing device can accurately predict the user's electricity consumption behavior.

[0076] Combination Figures 3a to 4c This paper describes the application of the environmentally adaptive power consumption behavior enhancement intelligent perception method based on multi-task learning to model training and multi-scale classification under multi-load and multi-switching power consumption behavior conditions. Specific details are as follows:

[0077] To obtain training features for identifying electricity consumption behavior categories, a small electric heater, a space heater, an induction cooker, a vacuum cleaner, and a water dispenser were connected to sockets. An oscilloscope was connected to the bus and current data was collected, with a sampling frequency of [missing information]. fs ( fs=50kHz) to obtain the current waveforms of five different electrical loads at the start and end of a fixed-length time period. Among them, the small electric heater uses resistance wire heating, and the induction cooker uses eddy currents generated by the magnetic field of a high-frequency electromagnetic field passing through a metal container, which then heats the metal container (resistor). Therefore, both the small electric heater and the induction cooker are resistance heating loads; while the space heater, vacuum cleaner, and water dispenser are non-resistance heating loads. For example... Figure 3a As shown, by artificially arranging the loads in a random order, the switching sequence of five types of electrical loads was determined (in order: small electric heater, induction cooker, electric heater, vacuum cleaner, water dispenser; this order is only for ease of understanding), thus forming the original current waveforms of resistive heating loads and non-resistive heating loads operating in different time domains.

[0078] The obtained current data (referring to) Figure 3a The current data corresponding to all waveforms are subjected to a short-time Fourier transform with a window length of 1024 to obtain the time-frequency distribution of the current data. Then, the frequency bands containing the current data features are filtered. Taking the features in the randomly selected 8.4kHz frequency band as an example (e.g.) Figure 3b As shown), the average time-frequency information (referred to as characteristic mean) of the five types of electrical loads is as follows: Small Sun 2.9×10 -4 A, Induction cooker 3.0×10 -4 A, Heater 6.7×10 -4 A, Vacuum cleaner 6.9×10 -4 A, Water dispenser 6.7×10 -4 A (according to) Figure 3b The characteristic mean of the corresponding load current data was calculated within the selected time-frequency interval. It can be seen that the characteristic mean of the small electric heater and the induction cooker are similar (characteristic error of 3.3%), while the characteristic mean of the heater, vacuum cleaner, and water dispenser are similar (with the highest characteristic error of 2.9%) and significantly different from the former (with the lowest characteristic error of 56.5%), making them clearly distinguishable. Therefore, the time-frequency information obtained in the 8.4kHz frequency band after the above short-time Fourier transform can meet the training features for identifying electricity consumption behavior categories. Furthermore, from... Figure 3b As can be seen, non-resistive heating loads and resistive heating loads have significantly different characteristics.

[0079] Continuing to randomly select other frequency bands, and setting the load status labels as 1 and 2 for two main load categories (resistive heating loads and non-resistive heating loads), a feature set for electricity consumption behavior category identification is formed. This feature set is then imported into the aforementioned multi-task learning improved temporal convolutional network for training. When the loss value is below 5%, training is considered complete, and the network can be used for electricity consumption behavior category identification. For example, when testing electricity consumption behavior category identification after training using feature combinations from the 5.1kHz, 5.8kHz, 6.8kHz, 7.1kHz, 7.2kHz, 7.6kHz, 7.7kHz, 8.4kHz, 9.3kHz, 10kHz, 10.9kHz, 11.7kHz, 13.6kHz, 15.5kHz, 19kHz, 21.8kHz, 24.7kHz, and 24.9kHz frequency bands, all features of the obtained current data are input into the trained multi-task learning improved temporal convolutional network. Based on the load category identification results output by this network, such as... Figure 3c As shown, the two types of labels corresponding to the two types of loads were almost perfectly identified (that is, the five types of electrical loads were classified into non-resistive heating loads and resistive heating loads), with a loss value of 4.61% and an accuracy rate of 99.63%.

[0080] Therefore, the electricity consumption behavior category identification proposed in this invention is a broad category identification. That is, when there is no need to perform specific and detailed load type classification, but only to care about the common differences between certain loads and other loads, it can accurately classify various loads into two major categories, greatly reducing calculation time and cost.

[0081] To obtain detailed training features for different types of electricity consumption behavior, the same set of current data was rearranged in different orders and then recombined to form the original current waveforms of five different types of loads operating in different time domains (in order: small electric heater, electric heater, induction cooker, vacuum cleaner, and water dispenser, respectively corresponding to...). Figure 4a The first, second, third, fourth, and fifth categories of loads.

[0082] The obtained current data (referring to) Figure 4a The current data corresponding to all waveforms are subjected to a short-time Fourier transform with a window length of 1024 to obtain the time-frequency distribution of the current data. Then, the frequency bands where the current data features are located are filtered. Taking the features under the randomly selected 194Hz frequency band as an example (e.g.) Figure 4b As shown), the average time-frequency information (referred to as characteristic mean) of the five types of electrical loads is as follows: Small Sun 2.3×10 -3 A, Heater 6.3×10 -3 A, Induction cooker 2.3×10 -2 A, Vacuum cleaner 5.4×10 -2A, Water dispenser 2.0×10 -3 A. It can be seen that the mini-solar heater and water dispenser have significant differences in characteristic mean compared to the heater, induction cooker, and vacuum cleaner (the lowest characteristic error is 63.4%). There are also significant differences in characteristic mean between each pair of heaters, induction cookers, and vacuum cleaners (the lowest characteristic error is 57.4%). However, the difference in characteristic mean between the mini-solar heater and water dispenser is not significant. Therefore, the time-frequency information of these two loads in the two time domain segments of 3.640s to 4.000s and 4.020s to 4.480s are taken, and their variances are calculated separately to determine their characteristic changes. The calculation shows that the characteristic variance of the mini-solar heater in the time domain segment of 3.640s to 4.000s is 3.9 × 10⁻¹⁰. -7 A 2 The characteristic variance in the time domain from 4.020s to 4.480s is 5.0 × 10⁻⁶. -8 A 2 It can be seen that the characteristic variance changes significantly between two adjacent time domain segments, and also exhibits a periodic trend (the time domain segments were chosen based on the periodicity of variance). The characteristic variance of the water dispenser in the time domain segment from 3.640s to 4.000s is 6.0 × 10⁻⁶. -7 A 2 The characteristic variance in the time domain from 4.020s to 4.480s is 5.5 × 10⁻⁶. -7 A 2 It can be seen that the feature variance changes little within two adjacent time domain segments, and the feature variance changes irregularly, indicating that the small heater and the water dispenser can be clearly distinguished (the feature error of variance change is 85.3%). Therefore, the time-frequency information in the 194Hz frequency band obtained after the above short-time Fourier transform can meet the training features for subdividing electricity consumption behavior types.

[0083] Continuing to randomly select other frequency bands, and assigning load status labels of 1, 2, 3, 4, and 5 to five different load types, a feature set for segmenting electricity consumption behavior types is formed. This feature set is then imported into the improved temporal convolutional network for multi-task learning and training. Training is considered complete when the loss value is below 5%, and the network can then be used for segmenting electricity consumption behavior types. For example, when testing electricity consumption behavior type segmentation using feature combinations from the 48Hz, 97Hz, 146Hz, 194Hz, 243Hz, 293Hz, 341Hz, 389Hz, 438Hz, 536Hz, 584Hz, 974Hz, 1.6kHz, 4.2kHz, and 20.2kHz frequency bands, all features of the obtained current data are input into the trained improved temporal convolutional network. Based on the load category identification results output by this network, such as... Figure 4cAs shown, the five tags corresponding to the five types of electrical loads were identified almost perfectly, with a loss value of 0.15% and an accuracy rate of 100%.

[0084] by Figure 4a Taking the current data shown as an example, after training a multi-task learning improved temporal convolutional network with training features for electricity consumption behavior category identification and electricity consumption behavior type subdivision, the extracted features (all features in each frequency band of the extracted time-frequency distribution) are input into this multi-task learning improved temporal convolutional network. After 150 iterations of training, the output prediction results for electricity consumption behavior category identification and electricity consumption behavior type subdivision are as follows. Figure 5a and Figure 5b As shown. From Figure 5a As can be seen, two different types of loads (resistive heating loads and non-resistive heating loads) are distinguished. From Figure 5b As can be seen, five different types of loads were distinguished. The total loss of the output results was 0.90%, the loss of the electricity behavior category identification results was 0.80%, the loss of the electricity behavior type segmentation results was 0.09%, the accuracy of the electricity behavior category identification was 99.88%, and the accuracy of the electricity behavior type segmentation was 100%.

[0085] The features of this invention are as follows:

[0086] 1) The environmental adaptive power consumption behavior enhancement intelligent perception method proposed in this invention can introduce multi-scale power consumption behavior classification enhancement perception technology, such as category identification, on the basis of the existing technical framework that only perceives the type of power consumption behavior, so as to meet the classification differentiation needs of different power consumption environments; it introduces remote control and comparison verification technology for power consumption behavior prediction and power load switching, forming an on-demand prediction and intelligent control of power equipment switching, and collecting power equipment classification perception information and verifying the correctness of prediction switching results after switching is completed, forming a closed-loop two-way interactive operation mode without human intervention, which greatly adapts to the differentiated needs of different users and significantly increases the intelligent power consumption experience.

[0087] 2) The electricity consumption behavior category identification proposed in this invention is a large-scale electricity consumption behavior classification technology. When the electricity consumption scenario does not require fine-grained classification of individual electrical equipment types, this technology can accurately classify electrical equipment into corresponding load states based on the commonalities and differences between them, greatly reducing the difficulty of feature construction and classification feature and model calculations. Regarding the fault detection function of different electrical equipment, problematic electrical equipment can also be identified by changing feature parameters and status indicators, and timely remote control switching can be implemented, thereby improving the power protection capabilities on the user side.

[0088] 3) The electricity consumption behavior prediction proposed in this invention is achieved by accurately predicting the future development trend of electricity consumption behavior in the current scenario and combining it with remote control of the switching of related electrical equipment. This realizes two-way interaction between electricity consumption behavior control and perception, creating conditions for significantly improving the level of smart homes under artificial intelligence conditions. It solves the drawback of traditional smart homes that can only perceive the single-direction change in electricity consumption behavior information flow of human-controlled electrical equipment, and effectively reduces the safety risks in the operation of electrical equipment.

[0089] 4) The technique proposed in this invention, which compares and verifies the prediction results of electricity consumption behavior with the multi-scale classification results, can form an internal enhancement effect of the electricity consumption behavior perception method. That is, after remotely controlling the switching of electrical equipment based on the prediction results, the electricity consumption behavior prediction results are verified by using the electricity consumption behavior classification information provided by the feedback. If the verification is incorrect at any scale, it will be adaptively corrected and adjusted through iterative training, thereby improving the accuracy of remote control and the ability to adapt to the environment.

[0090] 5) This invention trains on a proposed improved temporal convolutional network for multi-task learning. Multiple tasks, including electricity behavior category identification, electricity behavior type segmentation, and electricity behavior prediction, are completed simultaneously. Reliable prediction of electricity behavior is achieved through iterative training based on the electricity behavior type segmentation. This saves time by reducing the execution time of multiple tasks. The algorithm framework, which shares model parameters, reduces the computational and storage costs of the classification model. Furthermore, multi-task learning can alleviate the overfitting problem of single-task classification models under small sample conditions. The data augmentation process between different tasks also helps improve the model's generalization ability and enhances the overall learning effect across multiple tasks.

Claims

1. A method for enhancing intelligent perception of environmentally adaptive power consumption behavior, characterized in that: This adaptive power consumption behavior enhancement intelligent sensing method includes the following steps: 1) The load bus output current signal is sampled according to the sampling frequency. fs Sampling was performed to obtain the first j Load bus current data for each analysis period x j ; 2) Transfer the load bus current data x j Perform time-frequency transformation to obtain the time-frequency distribution of the current data, extract the time-frequency information of each frequency band in the time-frequency distribution to obtain the input features; 3) The multi-task learning model of the fused deep learning neural network trained by the training performs load type identification and classification on the input features at multiple scales, thereby simultaneously obtaining the output results of electricity behavior category identification and electricity behavior type subdivision, as well as the output results of electricity behavior prediction at the corresponding scale. 4) According to the first j The electricity consumption behavior prediction results obtained in each analysis period are used to remotely control the corresponding load to switch on and off in the next analysis period. The electricity consumption behavior type subdivision results and electricity consumption behavior category identification results obtained in the previous analysis period are compared and verified respectively. In the multi-task learning model that integrates deep learning neural networks, the deep learning neural network adopts a temporal convolutional network; The multi-task learning model that integrates deep learning neural networks combines a Keras-Shared-Bottom model with a multi-task learning model by modifying the fully connected layers in the temporal convolutional network. The output of the multi-task learning model that integrates deep learning neural networks is represented as follows: In the formula, h t The input at time t, y k For the first k The output obtained from each task k =1,2…, n , n For the number of tasks, b n This represents the deviation vector for each task. w n This is the weight matrix corresponding to each task; The training of the multi-task learning model fused with deep learning neural networks includes the following steps: collecting current data of the load to be identified during operation according to the analysis period length, performing corresponding time-frequency transformation, filtering the frequency bands of the obtained time-frequency distribution, and selecting frequency bands according to the requirement that the time-frequency information of the selected frequency bands under each classification scale can obtain common power consumption behavior feature parameters under the same classification state and / or obtain significantly different power consumption behavior feature parameters under different classification states. The time-frequency information under this frequency band is used as the training feature for power consumption behavior category identification or power consumption behavior type subdivision, and imported into the multi-task learning model fused with deep learning neural networks. The acquisition of current data of the load to be identified during operation specifically includes the following steps: using a sampling frequency of... fs The current waveforms of electrical loads are obtained at the start and end of a fixed time period. The current waveforms of each electrical load are combined to obtain the original current waveforms of different types of loads operating in different time domains. The screening process specifically includes the following steps: performing a short-time Fourier transform on the current data of the load to be identified during operation, and then randomly selecting time-frequency information from one or more frequency bands from the obtained time-frequency distribution; The characteristic parameters are one or more of the mean, variance, and variance variation of time-frequency information under the selected frequency band.

2. The environmental adaptive power consumption behavior enhancement intelligent sensing method according to claim 1, characterized in that: The sampling frequency fs The frequency range is 25~60kHz, and the number of sampling points during the analysis period is ≥3000.

3. The environmental adaptive power consumption behavior enhancement intelligent sensing method according to claim 1, characterized in that: The time-frequency transformation employs a short-time Fourier transform with a rectangular window, where the window length is 512~2048.

4. The environmental adaptive power consumption behavior enhancement intelligent sensing method according to claim 1, characterized in that: In step 4, after the corresponding load is switched on and off in the next analysis period, if the electricity behavior type subdivision result or electricity behavior category identification result obtained in the analysis period is inconsistent with the electricity behavior prediction result obtained in the previous analysis period, that is, the electricity behavior prediction result is not verified correctly, then the electricity behavior prediction result obtained in the corresponding analysis period is changed and iterative training is performed until the electricity behavior prediction result is consistent with the electricity behavior type subdivision result and electricity behavior category identification result obtained in the prediction analysis period, that is, the electricity behavior prediction result is verified correctly.

5. An environmentally adaptive power consumption behavior enhancement intelligent sensing device, characterized in that: The adaptive power consumption behavior enhancement intelligent sensing device includes a load bus output current signal sampling module, a load bus current data feature analysis module, a multi-scale power consumption behavior classification and power consumption behavior prediction module, and a comparison and verification module. The load bus output current signal sampling module is used to sample the load bus output current signal according to the sampling frequency. fs Sampling is performed until load bus current data for one analysis period is obtained; The load bus current data feature analysis module is used to analyze the first... j Load bus current data for each analysis period x j Perform time-frequency transformation and extract the time-frequency information of each frequency band in the time-frequency distribution obtained by time-frequency transformation as input features; The multi-scale electricity consumption behavior classification and prediction module is used to identify and classify load types at multiple scales and predict electricity consumption behavior at corresponding scales using a multi-task learning model of a trained fusion deep learning neural network. The comparison and verification module is used to compare and verify based on the first j The predicted electricity consumption behavior obtained in each analysis period is used to remotely control the corresponding load to switch on and off in the next analysis period. The predicted electricity consumption behavior behavior obtained in the previous analysis period is compared and verified based on the detailed breakdown results of electricity consumption behavior behavior type and the identification results of electricity consumption behavior category obtained in the analysis period. In the multi-task learning model that integrates deep learning neural networks, the deep learning neural network adopts a temporal convolutional network; The multi-task learning model that integrates deep learning neural networks combines a Keras-Shared-Bottom model with a multi-task learning model by modifying the fully connected layers in the temporal convolutional network. The output of the multi-task learning model that integrates deep learning neural networks is represented as follows: In the formula, h t The input at time t, y k For the first k The output obtained from each task k =1,2…, n , n For the number of tasks, b n This represents the deviation vector for each task. w n This is the weight matrix corresponding to each task; The training of the multi-task learning model fused with deep learning neural networks includes the following steps: collecting current data of the load to be identified during operation according to the analysis period length, performing corresponding time-frequency transformation, filtering the frequency bands of the obtained time-frequency distribution, and selecting frequency bands according to the requirement that the time-frequency information of the selected frequency bands under each classification scale can obtain common power consumption behavior feature parameters under the same classification state and / or obtain significantly different power consumption behavior feature parameters under different classification states. The time-frequency information under this frequency band is used as the training feature for power consumption behavior category identification or power consumption behavior type subdivision, and imported into the multi-task learning model fused with deep learning neural networks. The acquisition of current data of the load to be identified during operation specifically includes the following steps: using a sampling frequency of... fs The current waveforms of electrical loads are obtained at the start and end of a fixed time period. The current waveforms of each electrical load are combined to obtain the original current waveforms of different types of loads operating in different time domains. The screening process specifically includes the following steps: performing a short-time Fourier transform on the current data of the load to be identified during operation, and then randomly selecting time-frequency information from one or more frequency bands from the obtained time-frequency distribution; The characteristic parameters are one or more of the mean, variance, and variance variation of time-frequency information under the selected frequency band.

Images