AI-based user energy usage portrait analysis method and system

Abstract

The present invention provides an AI-based user energy usage portrait analysis method and system. By combining template user energy usage behavior data and a sequence of supplementary energy usage portrait description vectors, the initialized user energy usage portrait analysis model is cyclically optimized, which not only improves the accuracy of user energy usage behavior analysis, but also realizes a deep portrait description of the target user's energy usage behavior. By calculating the characteristic distance between the target energy usage portrait description vector and the supplementary energy usage portrait description vector, and selecting the template energy usage portrait label corresponding to the minimum distance as the label of the target user, the present invention can efficiently and accurately identify and label the energy usage behavior characteristics of the target user. In addition, in each model parameter optimization stage, the energy usage portrait description error is determined by predicting the correlation between the energy usage portrait description vector set and the supplementary energy usage portrait description vector set, and the model parameters are updated accordingly, further improving the training efficiency and generalization ability of the model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-based user energy usage portrait analysis method and system. Background Art

[0002] In the field of green energy supply management, in-depth understanding and analysis of user energy consumption behavior is of great significance for improving energy utilization efficiency, optimizing energy allocation, and formulating personalized service strategies. Traditionally, the analysis of user energy consumption behavior mainly relies on statistical methods and simple data analysis techniques, which face many challenges in dealing with complex and changeable user energy consumption data.

[0003] On the one hand, traditional statistical methods often rely on a large amount of historical data, and reveal the overall trends and patterns of users' energy consumption behaviors by calculating statistical indicators such as averages and standard deviations. However, this method often ignores the differences between individual users and the dynamic changes in energy consumption behaviors, resulting in inaccurate and intricate analysis results.

[0004] On the other hand, with the continuous increase and diversification of user energy consumption data, simple data analysis technology can no longer meet the needs of in-depth analysis of user energy consumption behavior. These user energy consumption data contain rich user behavior information and preference characteristics, but how to effectively extract and use this information has become an urgent problem to be solved. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an AI-based user energy usage portrait analysis method, the method comprising:

[0006] Acquire a template user energy usage behavior data sequence and a supplementary energy usage portrait description vector sequence; wherein the supplementary energy usage portrait description vector sequence is generated by performing energy usage portrait description on each template user energy usage behavior data in the template user energy usage behavior data sequence according to a pre-trained supplementary portrait analysis model;

[0007] According to the template user energy usage behavior data sequence and the supplementary energy usage portrait description vector sequence, the initialized user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model;

[0008] Acquire target user energy consumption behavior data, and perform energy consumption profile description on the target user energy consumption behavior data according to the target user energy consumption profile analysis model to generate a target energy consumption profile description vector;

[0009] Determine the target feature distances between the target energy usage portrait description vector and each of the supplementary energy usage portrait description vectors included in the supplementary energy usage portrait description vector sequence, and use the template energy usage portrait label associated with the supplementary energy usage portrait description vector corresponding to the minimum target feature distance as the target energy usage portrait label of the target user energy usage behavior data;

[0010] Among them, each model parameter optimization stage includes:

[0011] Performing energy usage portrait description on the candidate template user energy usage behavior data set according to the user energy usage portrait analysis model to generate a set of predicted energy usage portrait description vectors; wherein the candidate template user energy usage behavior data set includes multiple template user energy usage behavior data in the template user energy usage behavior data sequence; the predicted energy usage portrait description vector set includes predicted energy usage portrait description vectors of the multiple template user energy usage behavior data;

[0012] Based on the generated set of predicted energy usage portrait description vectors and their corresponding set of supplementary energy usage portrait description vectors, the energy usage portrait description error of the user energy usage portrait analysis model is determined, and based on the generated energy usage portrait description error, the neuron weight parameters of the user energy usage portrait analysis model are updated; wherein the energy usage portrait description error is used to reflect: the correlation between the description knowledge space of each predicted energy usage portrait description vector in the set of predicted energy usage portrait description vectors and the description knowledge space of each supplementary energy usage portrait description vector in the set of supplementary energy usage portrait description vectors.

[0013] On the other hand, an embodiment of the present invention also provides an AI-based user energy usage portrait analysis system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0014] Based on the above aspects, the embodiment of the present application combines the template user energy usage behavior data and the supplementary energy usage portrait description vector sequence to cyclically optimize the initialized user energy usage portrait analysis model, thereby generating an accurate target user energy usage portrait analysis model, which not only improves the accuracy of the user energy usage behavior analysis, but also realizes the deep portrait description of the target user's energy usage behavior, and by calculating the characteristic distance between the target energy usage portrait description vector and the supplementary energy usage portrait description vector, and selecting the template energy usage portrait label corresponding to the minimum distance as the label of the target user, the present invention can efficiently and accurately identify and label the energy usage behavior characteristics of the target user. In addition, in each model parameter optimization stage, the energy usage portrait description error is determined by predicting the correlation between the energy usage portrait description vector set and the supplementary energy usage portrait description vector set, and the model parameters are updated accordingly, which further improves the training efficiency and generalization ability of the model, making the user energy usage portrait analysis more intelligent and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the execution flow of the AI-based user energy usage portrait analysis method provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the hardware architecture of the AI-based user energy usage profile analysis system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an AI-based user energy usage portrait analysis method provided by an embodiment of the present invention. The AI-based user energy usage portrait analysis method is introduced in detail below.

[0018] Step S110, obtaining a template user energy usage behavior data sequence and a supplementary energy usage portrait description vector sequence, wherein the supplementary energy usage portrait description vector sequence is generated by performing energy usage portrait description on each template user energy usage behavior data in the template user energy usage behavior data sequence according to a pre-trained supplementary portrait analysis model.

[0019] In this embodiment, in a green energy supply service system, there are many users, and the server will select part of the data of these users as template user energy consumption behavior data. These template users may come from different regions, different industries, different energy-consuming equipment combinations and other complex factors. For example, some template users are large shopping malls located in commercial centers, and their energy consumption behavior data include daily lighting electricity consumption, air conditioning electricity consumption, elevator electricity consumption, etc., and these data have different performances in different time periods (such as peak hours, non-peak hours, weekends, etc. on weekdays); some template users are small factories, and their energy consumption behavior data involves the power consumption of production equipment, the power consumption of lighting and ventilation equipment in the workshop, etc., and will be affected by factors such as production plans and seasonal orders. The server collects the detailed energy consumption behavior data of these template users, arranges them in a certain order, and forms a template user energy consumption behavior data sequence.

[0020] Then, the pre-trained supplementary portrait analysis model can be used to describe the energy consumption profile of each template user's energy consumption behavior data in the template user energy consumption behavior data sequence. This supplementary portrait analysis model is obtained through a large amount of data training, and can deeply explore the complex energy consumption characteristics behind the energy consumption behavior data. Taking the template users of large shopping malls as an example, the supplementary portrait analysis model can analyze the changes in its energy consumption patterns in different seasons and during different promotional activities, as well as its energy efficiency compared with other surrounding shopping malls and other complex characteristics, and convert these characteristics into vector form to generate a supplementary energy consumption portrait description vector. For small factory template users, the vector representation corresponding to complex factors such as the energy consumption fluctuations during the adjustment of their production processes and the impact of equipment aging on energy consumption can be analyzed. These supplementary energy consumption portrait description vectors are arranged in the order of template users to form a sequence of supplementary energy consumption portrait description vectors.

[0021] Step S120, based on the template user energy usage behavior data sequence and the supplementary energy usage portrait description vector sequence, the initialized user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model.

[0022] In this embodiment, the user energy usage profile analysis model is first initialized. The user energy usage profile analysis model may be a deep neural network model including multiple hidden layers and neurons. The server starts a process of cyclically optimizing the model parameters.

[0023] Assume that the template user energy consumption behavior data sequence carries multiple complex template energy consumption profile labels. For example, for template users in large shopping malls, their template energy consumption profile labels may include complex profile labels such as "multi-regional coordinated energy consumption mode during peak hours", "special energy consumption adjustment strategy during promotional activities", and "interaction of energy consumption equipment on different floors". For template users in small factories, there may be labels such as "dynamic adjustment of equipment combination energy consumption based on order fluctuations" and "association between environmental protection equipment operation and main production equipment energy consumption".

[0024] For each template energy usage portrait label, the server selects the supplementary energy usage portrait description vectors of multiple template user energy usage behavior data corresponding to the current template energy usage portrait label from the supplementary energy usage portrait description vector sequence for fusion. For example, for the template energy usage portrait label of "multi-region collaborative energy usage mode during peak hours" of a large shopping mall, the server will find the supplementary energy usage portrait description vectors of the large shopping mall template users related to this label, and then fuse these energy usage portrait description vectors through a specific fusion algorithm (such as weighted average, etc.) to generate the supplementary energy usage portrait label vector corresponding to this template energy usage portrait label. The same operation is also applied to other template energy usage portrait labels.

[0025] Then, the user energy usage profile analysis model is cyclically optimized based on the template user energy usage behavior data sequence, the supplementary energy usage profile description vector sequence, and the supplementary energy usage profile label vectors corresponding to the generated multiple template energy usage profile labels. In each cycle, the server selects a portion of data from the template user energy usage behavior data sequence as a candidate template user energy usage behavior data set, which includes energy usage behavior data of different types of template users (such as shopping malls of different sizes, factories of different production types, etc.).

[0026] Step S130, acquiring the target user's energy usage behavior data, and performing an energy usage profile description on the target user's energy usage behavior data according to the target user's energy usage profile analysis model, to generate a target energy usage profile description vector.

[0027] In this embodiment, after the training of the target user energy consumption profile analysis model is completed, the energy consumption behavior data of the target user can be obtained. The target user may be a large shopping mall or factory that has newly joined the energy management system, or an existing user whose energy consumption profile needs to be re-evaluated. For example, the target user is a large shopping mall that has just undergone a partial area renovation (such as adding a new entertainment facility area).

[0028] The server decomposes the target user's energy consumption behavior data according to predefined rules. This predefined rule may be set based on a time window (such as one data block per hour) or data size (such as one data block for every 1,000 data items). For each data block, the server extracts features and generates an initial feature set for each data block. Taking a large shopping mall as an example, the statistical features in the initial feature set may include average power consumption per hour, maximum power, minimum power, power standard deviation, etc.; semantic features may include behavioral periodicity characteristics of energy consumption caused by the flow of people in the mall (such as the periodicity of power consumption caused by the difference in personnel flow on weekends and weekdays) and behavioral habit characteristics (such as the habit of turning on the air conditioner at a fixed time every day).

[0029] Then, each data block and the corresponding initial feature set are input into the initial layer of the target user energy consumption profile analysis model. In the initial layer, the mapping operation is performed according to the predefined neuron connections and weights and then converted into an intermediate feature representation. This intermediate feature representation integrates the relationship between the initial features in the initial feature set. For example, the average power consumption is associated with the periodic characteristics of personnel flow, reflecting the impact pattern of personnel flow on power consumption.

[0030] Next, the intermediate feature representation is transferred and processed between multiple hidden layers in the target user energy consumption profile analysis model. As the intermediate feature representation propagates between different hidden layers, each hidden layer further fuses and abstracts the intermediate feature representation. For example, in a hidden layer, the energy consumption features under different time windows are fused to obtain a more macro energy consumption trend feature; between adjacent hidden layers, the connection weights between neurons perform a weighted combination of the input intermediate feature representations to generate a high-level feature representation.

[0031] In the output layer of the target user energy consumption profile analysis model, a target energy consumption profile feature set related to the energy consumption profile is screened and combined from the high-level feature representation based on the pre-learned energy consumption profile construction mode. This target energy consumption profile feature set may include the time distribution pattern of the mall's energy consumption (such as the proportion of electricity consumption in different time periods), the total amount of energy consumption (such as the total electricity consumption per day, week, and month), and the stability of energy consumption (such as the fluctuation range of electricity power).

[0032] Finally, the initially constructed target energy consumption profile feature set is sorted and encoded to generate the target energy consumption profile description vector. For numerical energy consumption profile features (such as total electricity consumption), their values ​​are directly mapped to the corresponding dimensions of the target energy consumption profile description vector; for categorical energy consumption profile features (such as energy consumption type classification in different areas of the mall), they are converted into vector form using the unique hot encoding method.

[0033] Step S140, determine the target feature distances between the target energy usage portrait description vector and each supplementary energy usage portrait description vector contained in the supplementary energy usage portrait description vector sequence, and use the template energy usage portrait label associated with the supplementary energy usage portrait description vector corresponding to the smallest target feature distance as the target energy usage portrait label of the target user's energy usage behavior data.

[0034] In this embodiment, the target feature distance between the target energy consumption portrait description vector and each supplementary energy consumption portrait description vector in the supplementary energy consumption portrait description vector sequence can be calculated. Taking the target user of a large shopping mall as an example, for each vector in the supplementary energy consumption portrait description vector sequence (these vectors may come from supplementary portraits of different types of shopping malls or factories), the server calculates the distance with the target energy consumption portrait description vector in various dimensions (such as total energy consumption, energy consumption time distribution, etc.).

[0035] Then, the server finds the supplementary energy consumption portrait description vector corresponding to the minimum target feature distance. Assume that this supplementary energy consumption portrait description vector is a supplementary portrait from a large shopping mall that has many similarities with the target shopping mall in terms of operation mode and energy consumption equipment configuration. This supplementary energy consumption portrait description vector is associated with a complex template energy consumption portrait label, such as "multi-region collaborative optimization of energy consumption with energy-saving strategy adjustment", which reflects the energy consumption mode and management strategy of this similar shopping mall. The server uses this template energy consumption portrait label as the target energy consumption portrait label for the energy consumption behavior data of the target user (i.e., the target shopping mall), thereby completing the accurate classification and identification of the target user's energy consumption portrait.

[0036] Among them, each model parameter optimization stage includes:

[0037] Step S121, performing energy usage portrait description on the candidate template user energy usage behavior data set according to the user energy usage portrait analysis model to generate a set of predicted energy usage portrait description vectors. The candidate template user energy usage behavior data set includes multiple template user energy usage behavior data in the template user energy usage behavior data sequence. The predicted energy usage portrait description vector set includes predicted energy usage portrait description vectors of the multiple template user energy usage behavior data.

[0038] In this embodiment, the server inputs each data in the candidate template user energy consumption behavior data set into the user energy consumption portrait analysis model. Taking a large shopping mall as an example, if there is a large shopping mall energy consumption behavior data during a certain festival in the candidate template user energy consumption behavior data set, the model will describe the energy consumption portrait of this shopping mall during the festival according to the previously learned patterns and parameters. This description may involve complex factors such as the relationship between energy consumption in various areas (such as dining areas, shopping areas, entertainment areas, etc.) in the shopping mall during the festival, the adjustment range of energy consumption compared with usual times, and the impact of special activities (such as fireworks shows, etc.) on energy consumption, and convert these factors into a predicted energy consumption portrait description vector. Similarly, for the energy consumption behavior data of a small factory, the model will analyze the complex contents such as the energy consumption coordination of equipment under a specific production task, the chain reaction of raw material supply changes on energy consumption, and generate a corresponding predicted energy consumption portrait description vector. These predicted energy consumption portrait description vectors for different template users in the candidate template user energy consumption behavior data set constitute a predicted energy consumption portrait description vector set.

[0039] Step S122, based on the generated predicted energy usage portrait description vector set and its corresponding supplementary energy usage portrait description vector set, determine the energy usage portrait description error of the user energy usage portrait analysis model, and update the neuron weight parameter of the user energy usage portrait analysis model based on the generated energy usage portrait description error. The energy usage portrait description error is used to reflect the correlation between the description knowledge space of each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set and the description knowledge space of each supplementary energy usage portrait description vector in the supplementary energy usage portrait description vector set.

[0040] For each predicted energy consumption portrait description vector in the predicted energy consumption portrait description vector set and the corresponding vector in the supplementary energy consumption portrait description vector set, the characteristic distance between them can be determined. For example, for the predicted energy consumption portrait description vector of a large shopping mall during the holiday season and the corresponding supplementary energy consumption portrait description vector, the server can calculate their distances in various dimensions (such as energy consumption proportions in different regions, differences in peak energy consumption times, etc.). Then, for each predicted energy consumption portrait description vector, the server calculates the characteristic distance with other vectors in the supplementary energy consumption portrait description vector set except for its own corresponding vector, and performs fusion calculations. For example, for the predicted vector of a large shopping mall, after calculating the distance with the supplementary vectors of other shopping malls or factories, a fusion characteristic distance is generated through a specific fusion algorithm (such as summing and averaging, etc.).

[0041] Next, the characteristic distance between the predicted energy consumption portrait description vector and its corresponding supplementary energy consumption portrait description vector, as well as the proportional weight between the fusion characteristic distances corresponding to the predicted energy consumption portrait description vector, are determined, thereby determining the description knowledge space loss corresponding to the predicted energy consumption portrait description vector. For example, if the distance between the predicted vector and the corresponding supplementary vector on the key energy consumption indicator is large, and the fusion distance is relatively small, a larger description knowledge space loss is calculated according to the set weight.

[0042] The energy consumption profile description error of the user energy consumption profile analysis model is determined based on the sum of the description knowledge space losses corresponding to each predicted energy consumption profile description vector. If this error is large, it means that there is a large gap between the model's prediction and the actual supplementary profile.

[0043] Assume that the template energy usage portrait labels in the template user energy usage behavior data sequence are very complex as described above, and each template energy usage portrait label corresponds to a supplementary energy usage portrait label vector. The server performs array dot multiplication calculations on the arrays consisting of the supplementary energy usage portrait label vectors corresponding to these multiple template energy usage portrait labels, and each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set. For example, for the relevant arrays and vectors of a large shopping mall, the confidence that the template user energy usage behavior data corresponding to the predicted energy usage portrait description vector matches each template energy usage portrait label can be obtained through dot multiplication calculations.

[0044] Based on these confidences, the predicted energy usage profile labels of the template user energy usage behavior data corresponding to the predicted energy usage profile description vectors are determined. Then, based on the energy usage profile label loss between the predicted energy usage profile labels corresponding to each predicted energy usage profile description vector and its corresponding template energy usage profile label, the profile label prediction error of the user energy usage profile analysis model is determined.

[0045] Finally, the fusion calculation results of the energy usage profile description error and the profile label prediction error are determined to generate the model learning error of the user energy usage profile analysis model. If the model learning error does not meet the error convergence requirements (such as the error is greater than the preset threshold), the server will update the neuron weight parameters of the user energy usage profile analysis model. For example, using methods such as the gradient descent algorithm, the connection weights between neurons are adjusted according to the size and direction of the error, so that the model can better predict the energy usage profile in the next iteration.

[0046] Based on the above steps, the embodiment of the present application combines the template user energy usage behavior data and the supplementary energy usage portrait description vector sequence to perform cyclic optimization on the initialized user energy usage portrait analysis model, thereby generating an accurate target user energy usage portrait analysis model, which not only improves the accuracy of the user energy usage behavior analysis, but also realizes the deep portrait description of the target user's energy usage behavior, and by calculating the characteristic distance between the target energy usage portrait description vector and the supplementary energy usage portrait description vector, and selecting the template energy usage portrait label corresponding to the minimum distance as the label of the target user, the present invention can efficiently and accurately identify and label the energy usage behavior characteristics of the target user. In addition, in each model parameter optimization stage, the energy usage portrait description error is determined by predicting the correlation between the energy usage portrait description vector set and the supplementary energy usage portrait description vector set, and the model parameters are updated accordingly, which further improves the training efficiency and generalization ability of the model, making the user energy usage portrait analysis more intelligent and reliable.

[0047] In a possible implementation, the template user energy usage behavior data sequence carries a plurality of template energy usage portrait tags, and each template energy usage portrait tag carries a plurality of template user energy usage behavior data in the template user energy usage behavior data sequence.

[0048] Step S120 further includes:

[0049] For the multiple template energy usage portrait labels, the supplementary energy usage portrait description vectors of the multiple template user energy usage behavior data corresponding to the current template energy usage portrait label are selected from the supplementary energy usage portrait description vector sequence for fusion to generate a supplementary energy usage portrait label vector corresponding to the current template energy usage portrait label.

[0050] Based on the template user energy usage behavior data sequence, the supplementary energy usage portrait description vector sequence and the supplementary energy usage portrait label vectors corresponding to the multiple generated template energy usage portrait labels, the user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model.

[0051] In this embodiment, when constructing the target user energy consumption profile analysis model, the template user energy consumption behavior data sequence and the operations related to the multiple template energy consumption profile tags it carries are first processed. In this energy management system, the template user energy consumption behavior data sequence covers the energy consumption data of many different types of users, and each template energy consumption profile tag is associated with multiple template user energy consumption behavior data in the data sequence.

[0052] Taking large shopping malls and small factories as examples, the template energy use profile label of large shopping malls may contain complex labels such as "lighting and air conditioning coordinated energy use based on business hours adjustment", and small factories may have labels such as "equipment combination energy use strategy adjusted according to order batches". For each template energy use profile label, the server performs specific operations from the supplementary energy use profile description vector sequence. For example, for the template energy use profile label "lighting and air conditioning coordinated energy use based on business hours adjustment" of large shopping malls, the server will find the supplementary energy use profile description vectors of the energy use behavior data of multiple large shopping mall template users corresponding to this label in the supplementary energy use profile description vector sequence. When the business hours of these large shopping malls are adjusted, the energy use of lighting and air conditioning will be recorded in detail in their energy use behavior data, and the supplementary energy use profile description vectors will characterize this energy use situation from many aspects, such as the power change ratio of lighting and air conditioning at different business hours, the energy consumption correlation between the two, etc. The server fuses these supplementary energy use profile description vectors found through a specific fusion algorithm, such as a weighted fusion algorithm. During weighted fusion, different weights are assigned according to the importance of the energy consumption behavior data represented by different vectors. For example, a higher weight is assigned to the vector that reflects the close correlation between lighting and air-conditioning energy consumption, thereby generating a supplementary energy consumption portrait label vector corresponding to this template energy consumption portrait label.

[0053] After completing the generation of the supplementary energy portrait label vectors corresponding to all template energy portrait labels, the server begins to optimize the model parameters of the user energy portrait analysis model in a loop based on the template user energy behavior data sequence, the supplementary energy portrait description vector sequence, and these newly generated supplementary energy portrait label vectors. During the loop optimization process, the server selects different subsets from the template user energy behavior data sequence as input data. For large shopping malls, energy behavior data subsets of large shopping malls of different sizes and geographical locations can be selected. For small factories, energy behavior data subsets of small factories with different production types and different equipment configurations will also be selected. These data subsets are input into the user energy portrait analysis model, and the model processes these data according to the current parameter settings. At the same time, the supplementary energy portrait description vector sequence and the supplementary energy portrait label vector are also involved in the optimization process of the model. When processing these data, the model will continuously adjust its own parameters to better adapt to the data characteristics. For example, for the energy consumption behavior data of large shopping malls, the model can learn the complex patterns of lighting and air-conditioning energy consumption in different shopping malls in different seasons and during different promotional activities, and adjust the parameters according to the relevant vectors and supplementary energy portrait label vectors in the supplementary energy portrait description vector sequence, so that the model can accurately describe this energy consumption pattern. Similarly, for small factories, the model will learn the dynamic adjustment mode of equipment energy consumption under different order volumes, and optimize the parameters according to the relevant vectors and vectors. After multiple such cycles of optimization, until the model reaches the preset optimization standard, the server finally generates a pre-trained target user energy consumption portrait analysis model. This target user energy consumption portrait analysis model can accurately profile the energy consumption behavior of different users, whether they are large shopping malls, small factories or other types of users, providing a reliable foundation for subsequent target user energy consumption portrait analysis.

[0054] In a possible implementation, step S122 includes:

[0055] Step S1221, determining the characteristic distance between each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set and each supplementary energy usage portrait description vector in the supplementary energy usage portrait description vector set.

[0056] Step S1222: determining the energy usage profile description error of the user energy usage profile analysis model based on the generated feature distances.

[0057] In a possible implementation, step S1222 includes:

[0058] Step S1222-1, respectively, performs fusion calculation on the feature distances between each predicted energy portrait description vector in the predicted energy portrait description vector set and other supplementary energy portrait description vectors in the supplementary energy portrait description vector set except the supplementary energy portrait description vector corresponding to the predicted energy portrait description vector, to generate a fused feature distance corresponding to the predicted energy portrait description vector.

[0059] Step S1222-2, determine the characteristic distance between the predicted energy usage portrait description vector and its corresponding supplementary energy usage portrait description vector, and the proportional weight between the fusion characteristic distances corresponding to the predicted energy usage portrait description vector, and determine the description knowledge space loss corresponding to the predicted energy usage portrait description vector.

[0060] Step S1222-3, determining the energy usage portrait description error of the user energy usage portrait analysis model according to the sum of the description knowledge space losses corresponding to each of the generated predicted energy usage portrait description vectors.

[0061] In this embodiment, taking a large shopping mall as an example, in the set of predicted energy consumption portrait description vectors, there are predicted energy consumption portrait description vectors for different large shopping malls in specific time periods (such as holiday promotions). These predicted energy consumption portrait description vectors contain energy consumption forecasts for various areas of the shopping mall (such as shopping areas, dining areas, entertainment areas, etc.). The vectors in the set of supplementary energy consumption portrait description vectors are accurate energy consumption portrait description vectors obtained by processing actual data through the supplementary portrait analysis model. For each predicted energy consumption portrait description vector, the server calculates its characteristic distance with each supplementary energy consumption portrait description vector in the set of supplementary energy consumption portrait description vectors in each dimension, and then combines the distance results of each dimension to obtain the overall characteristic distance between the two vectors. In this way, this characteristic distance calculation is performed on each vector in the set of predicted energy consumption portrait description vectors and all vectors in the set of supplementary energy consumption portrait description vectors.

[0062] Then, the characteristic distances between each predicted energy use portrait description vector in the predicted energy use portrait description vector set and other supplementary energy use portrait description vectors in the supplementary energy use portrait description vector set except the supplementary energy use portrait description vector corresponding to the predicted energy use portrait description vector are fused and calculated to generate the fused characteristic distance corresponding to the predicted energy use portrait description vector. Taking a large shopping mall as an example, for a predicted energy use portrait description vector of a large shopping mall during a holiday promotion period, when calculating the characteristic distance fusion with other supplementary energy use portrait description vectors (except the corresponding supplementary vector that accurately describes the energy use of the mall during the promotion period), the server can adopt a weighted summation method. Assuming that there are multiple other supplementary vectors, which come from different large shopping malls or different time periods of the same shopping mall, the supplementary vectors of shopping malls with similar scale and business model to the shopping mall are given a higher weight in the fusion calculation, because they may be more referenced in energy use behavior. The fused characteristic distance corresponding to this predicted energy use portrait description vector is obtained through such a weighted summation calculation.

[0063] Then, the server determines the characteristic distance between the predicted energy consumption portrait description vector and its corresponding supplementary energy consumption portrait description vector, as well as the proportional weight between the fusion characteristic distances corresponding to the predicted energy consumption portrait description vector, so as to determine the description knowledge space loss corresponding to the predicted energy consumption portrait description vector. For the example of a large shopping mall, the characteristic distance between the predicted energy consumption portrait description vector and its corresponding supplementary energy consumption portrait description vector (i.e., the supplementary vector that accurately describes the energy consumption during the promotion period of the shopping mall) reflects the direct difference between the prediction and the actual. Assume that this direct characteristic distance is D1, and the fusion characteristic distance calculated above is D2. The server determines the proportional weight between them according to predefined rules, and this proportional weight may be set based on factors such as the reliability of the data and the importance of the vector. For example, if the direct characteristic distance D1 is considered to better reflect the accuracy of the model in a specific scenario, it can be given a higher weight. Then, the description knowledge space loss corresponding to the predicted energy consumption portrait description vector is calculated by a specific formula (such as a loss function formula). This loss reflects the impact of the description error of the model on this prediction vector on the overall knowledge space.

[0064] Finally, the server determines the energy use portrait description error of the user energy use portrait analysis model based on the sum of the description knowledge space losses corresponding to each generated prediction energy use portrait description vector. For each vector in the entire prediction energy use portrait description vector set, its description knowledge space loss is calculated according to the previous steps, and then these loss values ​​are summed. The sum result is the energy use portrait description error of the user energy use portrait analysis model. This error value comprehensively reflects the overall deviation between the model and the actual supplementary energy use portrait description vector when predicting the template user energy use behavior data. If this error value is large, it means that the prediction accuracy of the model is low, and it is necessary to further adjust the model parameters, such as neuron weights, to improve the performance of the model so that it can more accurately describe the user energy use portrait. For other types of users such as small factories, the server will also perform the same calculation process to comprehensively evaluate the accuracy of the model in describing the energy use portraits of different types of users. For example, for the energy use portrait prediction of a small factory under different production order quantities, the server calculates various distances between each prediction vector and the supplementary vector, fusion distance, determines the proportional weight, and calculates the description knowledge space loss, and finally sums up to obtain the error of the model in describing the energy use portrait of the small factory. Through such comprehensive and detailed calculations, the server can accurately evaluate the performance of the user energy consumption profile analysis model and provide a basis for subsequent model optimization.

[0065] In a possible implementation, the template user energy usage behavior data sequence carries multiple template energy usage portrait labels, each template energy usage portrait label carries multiple template user energy usage behavior data in the template user energy usage behavior data sequence, each template energy usage portrait label corresponds to a supplementary energy usage portrait label vector, and the supplementary energy usage portrait label vector is generated by fusing the supplementary energy usage portrait description vectors of the multiple template user energy usage behavior data corresponding to the corresponding template energy usage portrait label.

[0066] Step S122 includes:

[0067] Perform array dot multiplication on the array composed of the supplementary energy portrait label vectors corresponding to the multiple template energy portrait labels and each predicted energy portrait description vector in the predicted energy portrait description vector set to generate the confidence level of the template user energy usage behavior data corresponding to the predicted energy portrait description vector matching each template energy portrait label in the multiple template energy portrait labels.

[0068] Based on the generated confidence levels, the predicted energy usage portrait label of the template user energy usage behavior data corresponding to the predicted energy usage portrait description vector is determined.

[0069] The portrait label prediction error of the user energy usage portrait analysis model is determined based on the energy usage portrait label loss between the predicted energy usage portrait labels corresponding to each predicted energy usage portrait description vector and its corresponding template energy usage portrait label.

[0070] The neuron weight parameters of the user energy usage portrait analysis model are updated according to the energy usage portrait description error and the generated portrait label prediction error.

[0071] In a possible implementation manner, updating the neuron weight parameters of the user energy usage portrait analysis model according to the energy usage portrait description error and the generated portrait label prediction error includes:

[0072] Determine the fusion calculation result of the energy usage portrait description error and the portrait label prediction error, and generate the model learning error of the user energy usage portrait analysis model.

[0073] When it is determined that the generated model learning error does not meet the error convergence requirement, the neuron weight parameters of the user energy portrait analysis model are updated.

[0074] In this embodiment, first, array dot multiplication calculation is performed on the arrays consisting of the supplementary energy use portrait label vectors corresponding to the multiple template energy use portrait labels, and each predicted energy use portrait description vector in the predicted energy use portrait description vector set. Taking a large shopping mall as an example, it is assumed that there are multiple template energy use portrait labels such as "multi-region collaborative energy use mode during peak hours" and "special energy use adjustment strategy during promotion activities". Each label has a corresponding supplementary energy use portrait label vector, which is generated by fusing the supplementary energy use portrait description vectors of the energy use behavior data of multiple large shopping mall template users related to the label. For example, for the supplementary energy use portrait label vector corresponding to the "multi-region collaborative energy use mode during peak hours", it fuses the description vectors related to the energy use coordination of multiple large shopping malls in various areas (shopping areas, dining areas, entertainment areas, etc.) during peak hours. The predicted energy use portrait description vector set contains energy use portrait prediction vectors for different large shopping malls at different times. For each predicted energy use portrait description vector, the server performs dot multiplication calculation on the array consisting of it and the supplementary energy use portrait label vector. This dot multiplication calculation is based on the method of multiplying and summing the corresponding elements of the vector. Through such calculations, the confidence of the template user energy consumption behavior data corresponding to the predicted energy consumption portrait description vector matching each template energy consumption portrait label in multiple template energy consumption portrait labels is generated. For example, for the predicted energy consumption portrait description vector of a large shopping mall during a specific peak period, after the dot product calculation with the array, the confidence of the mall energy consumption behavior data matching the label "multi-regional coordinated energy consumption mode during peak period" is obtained as a certain value, and the confidence values ​​of matching other template energy consumption portrait labels (such as "special energy consumption adjustment strategy during promotion activities") are also obtained.

[0075] Next, based on the generated confidence levels, the server determines the predicted energy usage portrait label of the template user's energy usage behavior data corresponding to the predicted energy usage portrait description vector. The server will compare the confidence values ​​corresponding to each predicted energy usage portrait description vector. Still taking a large shopping mall as an example, if the predicted energy usage portrait description vector of a certain shopping mall is calculated with the confidence levels corresponding to each template energy usage portrait label, and it is found that its confidence level for the label "multi-region collaborative energy usage mode during peak hours" is the highest, then the server determines that the predicted energy usage portrait label of the template user (i.e., the large shopping mall) corresponding to this predicted energy usage portrait description vector is "multi-region collaborative energy usage mode during peak hours".

[0076] Then, according to the energy portrait label loss between the predicted energy portrait label corresponding to each predicted energy portrait description vector and its corresponding template energy portrait label, the portrait label prediction error of the user energy portrait analysis model is determined. For each predicted energy portrait description vector, the server compares its predicted energy portrait label with the corresponding template energy portrait label. For example, the actual template energy portrait label of a large shopping mall is "multi-regional collaborative energy consumption mode during peak hours", but the label predicted by the model is "conventional energy consumption mode during normal hours", which means there is a difference in labels. The server calculates the loss caused by this difference according to a predefined energy portrait label loss calculation method. This calculation method may take into account the degree of difference between labels in energy consumption mode, regional collaboration and other aspects, and quantify it into a numerical value. For all predicted energy portrait description vectors, the energy portrait label loss is calculated in this way, and then these loss values ​​are summed to obtain the portrait label prediction error of the user energy portrait analysis model. This error reflects the accuracy of the model in predicting energy portrait labels.

[0077] Afterwards, based on the energy portrait description error and the generated portrait label prediction error, the server updates the neuron weight parameters of the user energy portrait analysis model. First, determine the fusion calculation results of the energy portrait description error and the portrait label prediction error to generate the model learning error of the user energy portrait analysis model. This fusion calculation may be a simple weighted sum. For example, different weights are assigned according to the importance of the two errors to the overall accuracy of the model, and then the model learning error is calculated. For the energy portrait analysis of large shopping malls, if the energy portrait description error reflects the error in the numerical prediction of energy power, duration, etc. in each area of ​​the shopping mall, and the portrait label prediction error reflects the error in the label prediction of the overall energy consumption mode of the shopping mall, then the model learning error after the fusion calculation comprehensively reflects the overall deviation of the model in both numerical prediction and label prediction.

[0078] Finally, when the server determines that the generated model learning error does not meet the error convergence requirement, the neuron weight parameters of the user energy consumption profile analysis model are updated. The error convergence requirement may be a preset threshold or a convergence rule. For example, if the model learning error is greater than a preset threshold, it means that the accuracy of the model has not yet met the requirement. At this time, the server uses a specific weight update algorithm (such as a gradient descent-based algorithm) to update the neuron weight parameters of the user energy consumption profile analysis model. For the neuron weights in the energy consumption profile analysis model of large shopping malls, it may involve the weight adjustment of the relationship between energy consumption factors in different areas (such as lighting, air conditioning, elevators, etc.), as well as the adjustment of the weights of energy consumption influencing factors in different time periods (peak, normal, and valley periods). By adjusting these neuron weight parameters, the model can more accurately describe the energy consumption profile and predict the label when processing the same or similar large shopping mall energy consumption behavior data next time, thereby improving the overall performance of the model. For other types of users such as small factories, the server will also perform the same operation process, but there will be differences in specific energy consumption profile labels, energy consumption behavior data content, etc. For example, a small factory may involve energy consumption mode labels under different combinations of production equipment, such as "energy-saving energy consumption mode of equipment groups based on specific product production", etc. The server follows the same steps to calculate confidence, determine prediction labels, calculate label losses, fuse errors and update neuron weight parameters to ensure the accuracy and effectiveness of the model in energy consumption portrait analysis of small factories. The whole process is a comprehensive and meticulous process of optimizing model performance. By continuously adjusting the neuron weight parameters, the model can better adapt to the complex energy consumption behavior data of different types of users, thereby providing more accurate energy consumption portrait analysis results.

[0079] In a possible implementation, step S110 includes:

[0080] Step S111, obtaining a plurality of basic user energy usage behavior data.

[0081] Step S112: deriving features from the plurality of basic user energy usage behavior data respectively according to the feature derivation strategy, and generating at least one derived user energy usage behavior data associated with the corresponding basic user energy usage behavior data.

[0082] Step S113: constructing the template user energy usage behavior data sequence according to the plurality of basic user energy usage behavior data and the generated derivative user energy usage behavior data.

[0083] In a possible implementation, step S112 includes:

[0084] Step S1121, marking the characteristic distribution of the basic user energy usage behavior data, wherein the statistical characteristics of the basic user energy usage behavior data are determined from the perspective of the numerical distribution of the basic user energy usage behavior data. Starting from the time series characteristics of the basic user energy usage behavior data, the time-related characteristics of the basic user energy usage behavior data are analyzed, and the relationship between the category information in the basic user energy usage behavior data and the proportion of each category information in the basic user energy usage behavior data are determined.

[0085] Step S1122, based on the characteristic distribution of the basic user energy consumption behavior data, the policy parameters in the feature derivation strategy and the calling relationship and data flow path between different modules in the feature derivation strategy are adapted and adjusted to generate an adapted feature derivation strategy, wherein the adaptation and adjustment at least includes adding time-related calculation logic or adjusting parameters in the feature derivation strategy to adapt to the processing of time series data, or adjusting the module part related to category processing in the feature derivation strategy.

[0086] Step S1123, according to the grouping rules defined in the feature derivation strategy and the feature distribution of the basic user energy usage behavior data, the basic user energy usage behavior data is hierarchically grouped to generate multiple subsets with hierarchical relationships. During the grouping process, an association relationship index is established between each subset and the basic user energy usage behavior data and other subsets. The association relationship index records the position of the subset in the entire data system of the basic user energy usage behavior data and the logical relationship between the subset and other subsets. The basis for grouping includes: performing a first-level grouping according to the source of the basic user energy usage behavior data to distinguish data from different data sources, and then performing a second-level grouping within each first-level grouping according to the type of the basic user energy usage behavior data, and then performing a third-level grouping within each second-level grouping according to the time characteristics or energy usage device type of the basic user energy usage behavior data.

[0087] Step S1124: for each subset, according to the intra-group feature derivation rules in the adapted feature derivation strategy, basic feature derivation operations are performed using the relationship between the internal data of the subset to generate an intra-group basic derived feature set for each subset, wherein the intra-group feature derivation rules include basic mathematical operations on numerical data, for subsets with time series data, the difference and ratio between adjacent data points are calculated to reflect the changes in the data points over time, and for subsets with category information, basic category-related features are generated based on the relationship between the category information, and the basic category-related features include category proportion and category conversion frequency.

[0088] Step S1125 , analyzing the relationship between different subsets according to the inter-group relationship feature derivation rule in the adapted and adjusted feature derivation strategy to generate an inter-group relationship feature.

[0089] Step S1126, according to the multi-feature fusion derivation rules in the adapted and adjusted feature derivation strategy, the derived features of each subset are fused to generate a multi-feature fusion derived feature set for each subset, wherein the derived features include the basic derived features within the group and the relationship features between groups.

[0090] Step S1127, according to the feature optimization rules in the adapted and adjusted feature derivation strategy, an optimization operation is performed on the multi-feature fusion derived feature set of each subset, and the optimization operation includes feature normalization processing and feature smoothing processing.

[0091] Step S1128, according to the feature screening rules in the adapted feature derivation strategy, the optimized subset derived feature sets are screened, and after the optimized and screened subset derived feature sets are globally integrated, the globally integrated derived features are renumbered and indexed to generate at least one derived user energy usage behavior data.

[0092] In this embodiment, when obtaining the template user energy consumption behavior data sequence, it is necessary to first obtain multiple basic user energy consumption behavior data. In the scenario of the energy management system, these basic user energy consumption behavior data come from a wide range of sources. For example, basic users may include different types of energy-consuming units such as large shopping malls, small factories, and office buildings. For large shopping malls, their basic user energy consumption behavior data covers the changes in the power consumption of the lighting system over time, the energy consumption data of the air-conditioning system in different seasons and different business hours, the relationship between the number of elevator operations and power consumption, etc.; the basic user energy consumption behavior data of small factories includes the power consumption of production equipment under different production tasks, the energy consumption of workshop lighting and ventilation equipment, and the energy utilization efficiency in the raw material processing process; the basic user energy consumption behavior data of office buildings includes the power consumption of electrical equipment (computers, printers, etc.) in the office area, and the lighting and air-conditioning energy consumption in public areas (corridors, halls).

[0093] Next, feature derivation is performed on these multiple basic user energy usage behavior data according to a feature derivation strategy to generate at least one derived user energy usage behavior data associated with the corresponding basic user energy usage behavior data.

[0094] In this process, we must first mark the characteristic distribution of the basic user energy consumption behavior data. Determine its statistical characteristics from the perspective of the numerical distribution of the basic user energy consumption behavior data. Taking the lighting power consumption data of large shopping malls as an example, the statistical characteristics include the mean, standard deviation, maximum value, minimum value, etc. Starting from the characteristics of time series, analyze the time-related characteristics of the basic user energy consumption behavior data. For example, the air conditioning energy consumption data of large shopping malls shows different trends within 24 hours a day, and the patterns on weekdays and weekends may be different. Through analysis, we can determine its time-related characteristics such as peak hours, trough hours, and periodicity of energy consumption fluctuations. At the same time, determine the relationship between the category information in the basic user energy consumption behavior data and the proportion of each category information in the basic user energy consumption behavior data. For large shopping malls, the category information may include the energy consumption of different floors (such as shopping floors, catering floors, and entertainment floors), the proportion of energy consumption of each floor in the total energy consumption of the mall, and the correlation between energy consumption between different floors (such as the extension of the business hours of the entertainment floor may lead to an increase in the passenger flow of the adjacent shopping floor, thereby affecting the lighting and air conditioning energy consumption of the shopping floor).

[0095] According to the characteristic distribution of basic user energy consumption behavior data, the policy parameters in the feature derivation strategy and the calling relationship and data flow path between different modules in the feature derivation strategy are adapted and adjusted to generate an adapted feature derivation strategy. For data with obvious time series characteristics (such as daily electricity consumption data of large shopping malls), it may be necessary to increase the time-related calculation logic in the feature derivation strategy or adjust the parameters to adapt to the processing of time series data. For example, the original calculation may not take into account the impact of different seasons on electricity consumption. After adjustment, the weight of seasonal factors can be increased or seasonal time functions can be introduced to analyze the data more accurately. For data containing category information (such as energy consumption classification of different types of users), it may be necessary to adjust the module part related to category processing in the feature derivation strategy. For example, if the previous category processing did not take into account the energy consumption differences between users in different industries, different category processing parameters can be set for different industries after adjustment.

[0096] Then, according to the grouping rules defined in the feature derivation strategy, the basic user energy behavior data is hierarchically grouped according to the feature distribution of the basic user energy behavior data, and multiple subsets with hierarchical relationships are generated. Taking a large shopping mall as an example, the basic user energy behavior data is grouped according to the source of the first-level grouping, and the data from different meters (such as lighting meters, air conditioning meters, etc.) can be distinguished. Within each first-level grouping, the basic user energy behavior data is grouped according to the type. For example, in the lighting meter data, the second-level grouping can be performed according to the lighting area (indoor lighting, outdoor sign lighting, etc.). Then, within each second-level grouping, the basic user energy behavior data is grouped according to the time characteristics or energy-consuming equipment type. For example, indoor lighting data can be further grouped according to different time characteristics during the day and at night, or according to different types of lighting fixtures (incandescent lamps, energy-saving lamps, LED lamps). In the grouping process, an association relationship index is established between each subset and the basic user energy behavior data and other subsets. This association relationship index records the position of the subset in the entire data system of the basic user energy behavior data and the logical relationship between it and other subsets. For example, the indoor lighting subset has an inclusion relationship with the lighting power consumption data of the entire shopping mall (basic user energy consumption behavior data), and at the same time has a comparison or correlation relationship with the outdoor sign lighting subset in the overall logic of lighting power consumption (e.g., the peak power consumption of the two may be different, but both are affected by the business hours of the shopping mall).

[0097] For each subset, according to the intra-group feature derivation rules in the adapted feature derivation strategy, the relationship between the internal data of the subset is used to perform basic feature derivation operations to generate the intra-group basic derived feature set of each subset. For numerical data subsets (such as the lighting power data subset of a certain area in a large shopping mall), basic mathematical operations are performed, such as calculating statistical derived features such as averages and sums. For subsets with time series data (such as a data subset of air conditioning energy consumption in a shopping mall that changes over time), the difference and ratio between adjacent data points are calculated to reflect the changes in data points over time. For example, by calculating the difference in air conditioning energy consumption between adjacent hours, the increase and decrease trend of energy consumption can be obtained, and by calculating the ratio, the relative change ratio of energy consumption can be obtained. For subsets with category information (such as energy consumption data subsets of different functional areas in a shopping mall), basic category-related features are generated based on the relationship between category information, such as calculating the energy consumption proportion of different functional areas (such as the proportion of energy consumption in the shopping area to the total energy consumption of the shopping mall) and the category conversion frequency (such as the frequency of a certain area switching from a low energy consumption state to a high energy consumption state).

[0098] According to the inter-group relationship feature derivation rules in the adapted feature derivation strategy, the relationship between different subsets is analyzed to generate inter-group relationship features. For example, when analyzing the relationship between the lighting and air conditioning electronic sets in large shopping malls, it may be found that there is a certain time lag between the peak lighting power consumption period and the peak air conditioning power consumption period (because the air conditioning will be adjusted according to the indoor temperature some time after the lighting is turned on). This is an inter-group relationship feature.

[0099] According to the multi-feature fusion derivation rules in the adapted feature derivation strategy, the derived features of each subset (including the basic derived features within the group and the inter-group relationship features) are fused to generate the multi-feature fusion derived feature set of each subset. For the lighting power consumption data subset of large shopping malls, its statistical features (such as average power), time-related features (such as power changes in different time periods) and inter-group relationship features (such as the correlation with air-conditioning power consumption) with other subsets (such as the air-conditioning power consumption set) are fused to obtain a multi-feature fusion derived feature set that more comprehensively reflects the lighting power consumption.

[0100] According to the feature optimization rules in the adapted feature derivation strategy, the multi-feature fusion derived feature set of each subset is optimized, and the optimization operation includes feature standardization and feature smoothing. For the multi-feature fusion derived feature set of lighting power consumption in large shopping malls, feature standardization can adjust features of different magnitudes (such as the power value range may be large, while the correlation value with air-conditioning power consumption may be small) to the same magnitude range to facilitate subsequent analysis and calculation. Feature smoothing can smooth some data with large fluctuations. For example, lighting power may fluctuate greatly due to the instantaneous opening or closing of electrical appliances. More stable trend data can be obtained through smoothing.

[0101] According to the feature screening rules in the adapted feature derivation strategy, the optimized subset derived feature sets are screened. The basis for screening may be factors such as the importance and relevance of the features. For example, for the lighting power consumption data of a large shopping mall, some features that are less relevant to the overall energy consumption trend (such as extremely small power fluctuations at a certain moment) may be screened out. After the optimized and screened subset derived feature sets are globally integrated, the globally integrated derived features are renumbered and indexed to generate at least one derived user energy consumption behavior data. This derived user energy consumption behavior data integrates various useful information of the basic user energy consumption behavior data after a series of complex processing, thereby providing richer and more valuable data content for constructing a template user energy consumption behavior data sequence. Finally, based on multiple basic user energy consumption behavior data and the generated derived user energy consumption behavior data, a template user energy consumption behavior data sequence is constructed. This sequence contains various feature information derived from the original basic data, which can more comprehensively reflect the user's energy consumption behavior situation and provide a data basis for subsequent training and optimization of the user energy consumption portrait analysis model.

[0102] In a possible implementation, step S130 includes:

[0103] Step S131, decompose the target user's energy usage behavior data according to predefined rules to generate multiple data blocks, and perform feature extraction on each data block to generate an initial feature set for each data block. The predefined rules are set based on a time window or a data volume. The initial feature set includes statistical features and semantic features. The statistical features include average features, maximum features, minimum features, and standard deviation features. The semantic features include behavioral periodicity features and behavioral habit features of the user's energy usage behavior.

[0104] Step S132, input each data block and the corresponding initial feature set into the initial layer of the target user energy usage portrait analysis model, and convert them into intermediate feature representations after mapping operations based on the predefined neuron connections and weights of the initial layer. The intermediate feature representation is a feature representation converted in the feature direction used to describe the energy usage portrait label, integrating the relationship between the initial features in the initial feature set.

[0105] Step S133, the intermediate feature representation is transferred and processed between multiple hidden layers in the target user energy usage portrait analysis model, wherein, as the intermediate feature representation propagates between different hidden layers, each hidden layer further fuses and abstracts the intermediate feature representation, and between adjacent hidden layers, the connection weights between neurons perform weighted combination on the input intermediate feature representation to generate a high-level feature representation.

[0106] Step S134, in the output layer of the target user energy usage portrait analysis model, based on the pre-learned energy usage portrait construction pattern, a target energy usage portrait feature set related to the energy usage portrait is screened and combined from the high-level feature representation, and the target energy usage portrait feature set includes the time distribution pattern of energy usage, the total amount of energy usage characteristics and the stability characteristics of energy usage.

[0107] Step S135, organizes and encodes the initially constructed target energy usage portrait feature set to generate a target energy usage portrait description vector, wherein each dimension of the target energy usage portrait description vector corresponds to an energy usage portrait feature in the target energy usage portrait feature set, wherein for numerical energy usage portrait features, their numerical values ​​are directly mapped to the corresponding dimensions of the target energy usage portrait description vector, and for categorical energy usage portrait features, they are converted into vector form using one-hot encoding.

[0108] In this embodiment, it is still assumed that the target user is a large shopping mall, and its energy consumption behavior data includes the power consumption information of many devices and areas in the shopping mall, such as the power consumption, power consumption duration and other data of each device such as the lighting system, air conditioning system, and elevator system, and these data are recorded in chronological order. If the predefined rule is set based on the time window, for example, one hour is a time window, and the energy consumption data of 24 hours a day is divided into 24 data blocks; if it is set based on the size of the data, it is assumed that every 1000 power consumption data records are recorded as a data block. For each data block, feature extraction is performed to generate an initial feature set. Taking the lighting system power consumption data block as an example, the statistical features include the average value feature, that is, the average value of the lighting power consumption in the data block, reflecting the average level of lighting power consumption during this period; the maximum value feature shows the highest value of the lighting power consumption, which may correspond to the power consumption at certain special moments in the shopping mall (such as when all lighting is turned on during promotional activities); the minimum value feature reflects the lowest power of lighting power consumption, which may be the state after some lighting in the shopping mall is turned off at night; the standard deviation feature characterizes the fluctuation of the lighting power consumption in the data block. In terms of semantic features, the behavioral periodicity of user energy consumption behavior is manifested in the periodic changes in lighting electricity consumption within a day. For example, electricity consumption is higher during daytime business hours and lower after closing at night. This periodicity reflects the regularity of the shopping mall's business hours; behavioral habit characteristics may be reflected in the special pattern of lighting electricity consumption during specific activities in the shopping mall (such as fixed promotion days every week).

[0109] For each data block of a large shopping mall and its initial feature set, in the initial layer of the target user energy consumption profile analysis model, neurons map the input features according to predefined connections and weights. For example, for statistical features such as average value features and maximum value features in the lighting system data block, as well as semantic features such as behavioral periodicity features and behavioral habit features, the neurons in the initial layer combine and transform these features through specific weight assignments. This process integrates the relationship between the initial features in the initial feature set, so that the intermediate feature representation is transformed in the direction of features used to describe energy consumption profile labels. For example, the average value of lighting electricity consumption is associated with the periodic feature to form an intermediate feature representation that can reflect the comprehensive characteristics of lighting electricity consumption in time and value. This intermediate feature representation is no longer a simple enumeration of the original features, but a new feature representation after weighting and mapping by neurons.

[0110] Then, for the intermediate feature representation of energy consumption data of large shopping malls, in the first hidden layer, the intermediate feature representation of lighting power consumption can be fused with the intermediate feature representation of air conditioning power consumption. Considering the mutual influence relationship between lighting and air conditioning in the shopping mall environment, such as the heat generated by lighting may affect the cooling demand of air conditioning, this relationship is integrated into the new feature representation, which is a fusion operation; at the same time, the features are abstracted through the neurons in the hidden layer, ignoring some minor details and extracting more representative features. Between adjacent hidden layers, the connection weights between neurons perform weighted combination of the input intermediate feature representation to generate a high-level feature representation. For example, in the process of transferring from the first hidden layer to the second hidden layer, the fused and abstracted intermediate feature representation output by the first hidden layer is weighted combined according to the connection weight, so that a higher-level and more comprehensive high-level feature representation is obtained in the second hidden layer. This high-level feature representation not only contains the comprehensive information of the power consumption of each system such as lighting and air conditioning, but also reflects the complex relationship between them and the overall characteristics in time and space.

[0111] In the output layer of the target user energy consumption profile analysis model, the target energy consumption profile feature set related to the energy consumption profile is screened and combined from the high-level feature representation based on the pre-learned energy consumption profile construction mode. For the target user of large shopping malls, the pre-learned energy consumption profile construction mode may contain various knowledge and rules about the overall energy consumption of the shopping malls. In the output layer, the features related to the energy consumption profile are screened from the high-level feature representation. The target energy consumption profile feature set includes the time distribution pattern of energy consumption, such as the distribution of electricity consumption in different time periods of a day (such as morning, noon, afternoon, and evening) or different days of a week (weekdays and weekends); the total amount of energy consumption, that is, the total electricity consumption of the shopping mall in a specific time period (such as one day, one week, and one month). This feature can reflect the overall energy consumption scale of the shopping mall; the stability of energy consumption, which is reflected in the fluctuation range of the shopping mall's electricity power. If the fluctuation is small, it means that the energy consumption stability is high, otherwise it is low. These features describe the energy consumption profile of large shopping malls from different angles.

[0112] Finally, the initially constructed target energy consumption profile feature set is sorted and encoded to generate a target energy consumption profile description vector. For numerical energy consumption profile features, their values ​​are directly mapped to the corresponding dimensions of the target energy consumption profile description vector. For example, if the total energy consumption feature is a specific value (such as the total electricity consumption in a month is 10,000 kWh), this value is directly mapped to the dimension in the target energy consumption profile description vector that is specifically used to represent the total electricity consumption. For categorical energy consumption profile features, they are converted into vector form using a one-hot encoding method. Assuming that the energy consumption types of shopping malls are divided into three categories: high energy consumption (such as during promotional activities), medium energy consumption (normal business days) and low energy consumption (closed at night), after using one-hot encoding, if the shopping mall is currently in a medium energy consumption state, the corresponding dimension in the target energy consumption profile description vector is marked as 1, and the dimensions corresponding to the other two categories are marked as 0. In this way, the categorical energy consumption profile features are converted into vector form, thereby completing the generation of the target energy consumption profile description vector. This target energy consumption profile description vector comprehensively describes the energy consumption profile of the target user (large shopping mall), and provides an effective data representation for subsequent analysis (such as comparison and classification with other energy consumption profiles).

[0113] Figure 2 The hardware structure of the AI-based user energy usage profile analysis system 100 for implementing the above-mentioned AI-based user energy usage profile analysis method provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the AI-based user energy usage portrait analysis system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0114] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the AI-based user energy usage profile analysis system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0115] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AI-based user energy usage portrait analysis method of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0116] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned AI-based user energy usage portrait analysis system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0117] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned AI-based user energy usage portrait analysis method is implemented.

[0118] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A user energy usage profile analysis method based on AI, characterized in that: The method comprises: Acquire a template user energy usage behavior data sequence and a supplementary energy usage portrait description vector sequence; wherein the supplementary energy usage portrait description vector sequence is generated by performing energy usage portrait description on each template user energy usage behavior data in the template user energy usage behavior data sequence according to a pre-trained supplementary portrait analysis model; According to the template user energy usage behavior data sequence and the supplementary energy usage portrait description vector sequence, the initialized user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model; Acquire target user energy consumption behavior data, and perform energy consumption profile description on the target user energy consumption behavior data according to the target user energy consumption profile analysis model to generate a target energy consumption profile description vector; Determine the target feature distances between the target energy usage portrait description vector and each of the supplementary energy usage portrait description vectors included in the supplementary energy usage portrait description vector sequence, and use the template energy usage portrait label associated with the supplementary energy usage portrait description vector corresponding to the minimum target feature distance as the target energy usage portrait label of the target user energy usage behavior data; Among them, each model parameter optimization stage includes: Performing energy usage portrait description on the candidate template user energy usage behavior data set according to the user energy usage portrait analysis model to generate a set of predicted energy usage portrait description vectors; wherein the candidate template user energy usage behavior data set includes multiple template user energy usage behavior data in the template user energy usage behavior data sequence; the predicted energy usage portrait description vector set includes predicted energy usage portrait description vectors of the multiple template user energy usage behavior data; Based on the generated set of predicted energy usage portrait description vectors and their corresponding set of supplementary energy usage portrait description vectors, the energy usage portrait description error of the user energy usage portrait analysis model is determined, and based on the generated energy usage portrait description error, the neuron weight parameters of the user energy usage portrait analysis model are updated; wherein the energy usage portrait description error is used to reflect: the correlation between the description knowledge space of each predicted energy usage portrait description vector in the set of predicted energy usage portrait description vectors and the description knowledge space of each supplementary energy usage portrait description vector in the set of supplementary energy usage portrait description vectors.

2. The AI-based user energy usage profile analysis method according to claim 1 is characterized in that: The template user energy usage behavior data sequence carries a plurality of template energy usage portrait tags, and each template energy usage portrait tag carries a plurality of template user energy usage behavior data in the template user energy usage behavior data sequence; Then, based on the template user energy usage behavior data sequence and the supplementary energy usage portrait description vector sequence, the initialized user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model, including: For the multiple template energy usage portrait labels, respectively select the supplementary energy usage portrait description vectors of the multiple template user energy usage behavior data corresponding to the current template energy usage portrait label from the supplementary energy usage portrait description vector sequence, and fuse them to generate a supplementary energy usage portrait label vector corresponding to the current template energy usage portrait label; Based on the template user energy usage behavior data sequence, the supplementary energy usage portrait description vector sequence and the supplementary energy usage portrait label vectors corresponding to the multiple generated template energy usage portrait labels, the user energy usage portrait analysis model is cyclically optimized to generate a pre-trained target user energy usage portrait analysis model.

3. The user energy usage profile analysis method based on AI according to claim 1 is characterized in that: The step of determining the energy usage portrait description error of the user energy usage portrait analysis model based on the generated predicted energy usage portrait description vector set and its corresponding supplementary energy usage portrait description vector set includes: Determine a characteristic distance between each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set and each supplementary energy usage portrait description vector in the supplementary energy usage portrait description vector set; The energy usage profile description error of the user energy usage profile analysis model is determined based on the generated feature distances.

4. The AI-based user energy usage profile analysis method according to claim 3 is characterized in that: The step of determining the energy usage profile description error of the user energy usage profile analysis model based on the generated characteristic distances includes: Performing fusion calculation on the characteristic distances between each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set and other supplementary energy usage portrait description vectors in the supplementary energy usage portrait description vector set except the supplementary energy usage portrait description vector corresponding to the predicted energy usage portrait description vector, to generate a fusion characteristic distance corresponding to the predicted energy usage portrait description vector; Determine the characteristic distance between the predicted energy usage portrait description vector and its corresponding supplementary energy usage portrait description vector, and the proportional weight between the fusion characteristic distances corresponding to the predicted energy usage portrait description vector, and determine the description knowledge space loss corresponding to the predicted energy usage portrait description vector; The energy usage portrait description error of the user energy usage portrait analysis model is determined based on the sum of the description knowledge space losses corresponding to each of the generated predicted energy usage portrait description vectors.

5. The user energy usage profile analysis method based on AI according to claim 1 is characterized in that: The template user energy usage behavior data sequence carries multiple template energy usage portrait tags, each template energy usage portrait tag carries multiple template user energy usage behavior data in the template user energy usage behavior data sequence, each template energy usage portrait tag corresponds to a supplementary energy usage portrait label vector, and the supplementary energy usage portrait label vector is generated by fusing the supplementary energy usage portrait description vectors of multiple template user energy usage behavior data corresponding to the corresponding template energy usage portrait tag; Then, the neuron weight parameters of the user energy usage profile analysis model are updated based on the generated energy usage profile description error, including: Perform array dot multiplication calculations on the arrays of supplementary energy usage portrait label vectors corresponding to the multiple template energy usage portrait labels and each predicted energy usage portrait description vector in the predicted energy usage portrait description vector set, and generate a confidence level that the template user energy usage behavior data corresponding to the predicted energy usage portrait description vector matches each template energy usage portrait label in the multiple template energy usage portrait labels; Determining, based on each generated confidence level, a predicted energy usage portrait label of the template user energy usage behavior data corresponding to the predicted energy usage portrait description vector; Determining the portrait label prediction error of the user energy consumption portrait analysis model according to the energy consumption portrait label loss between the predicted energy consumption portrait labels corresponding to the predicted energy consumption portrait description vectors and their corresponding template energy consumption portrait labels; The neuron weight parameters of the user energy usage portrait analysis model are updated according to the energy usage portrait description error and the generated portrait label prediction error.

6. The AI-based user energy usage profile analysis method according to claim 5 is characterized in that: The updating of the neuron weight parameters of the user energy usage portrait analysis model according to the energy usage portrait description error and the generated portrait label prediction error includes: Determine a fusion calculation result of the energy usage portrait description error and the portrait label prediction error to generate a model learning error of the user energy usage portrait analysis model; When it is determined that the generated model learning error does not meet the error convergence requirement, the neuron weight parameters of the user energy portrait analysis model are updated.

7. The method for analyzing user energy usage profile based on AI according to any one of claims 1 to 6, characterized in that: The step of obtaining the template user energy usage behavior data sequence includes: Obtain multiple basic user energy consumption behavior data; According to the feature derivation strategy, feature derivation is performed on the plurality of basic user energy usage behavior data respectively to generate at least one derived user energy usage behavior data associated with the corresponding basic user energy usage behavior data; The template user energy usage behavior data sequence is constructed based on the multiple basic user energy usage behavior data and the generated derived user energy usage behavior data.

8. The AI-based user energy usage profile analysis method according to claim 7 is characterized in that: The step of performing feature derivation on the plurality of basic user energy usage behavior data respectively according to the feature derivation strategy to generate at least one derived user energy usage behavior data associated with the corresponding basic user energy usage behavior data comprises: Marking the characteristic distribution of the basic user energy usage behavior data, wherein the statistical characteristics of the basic user energy usage behavior data are determined from the perspective of the numerical distribution of the basic user energy usage behavior data; starting from the time series characteristics of the basic user energy usage behavior data, the time-related characteristics of the basic user energy usage behavior data are analyzed, and the relationship between the category information in the basic user energy usage behavior data and the proportion of each category information in the basic user energy usage behavior data are determined; According to the characteristic distribution of the basic user energy consumption behavior data, the policy parameters in the characteristic derivation strategy and the calling relationship and data flow path between different modules in the characteristic derivation strategy are adapted and adjusted to generate an adapted characteristic derivation strategy, wherein the adaptation and adjustment at least includes adding time-related calculation logic or adjusting parameters in the characteristic derivation strategy to adapt to the processing of time series data, or adjusting the module part related to category processing in the characteristic derivation strategy; According to the grouping rules defined in the feature derivation strategy and according to the feature distribution of the basic user energy usage behavior data, the basic user energy usage behavior data is hierarchically grouped to generate multiple subsets with hierarchical relationships. In the grouping process, an association relationship index is established between each subset and the basic user energy usage behavior data and other subsets. The association relationship index records the position of the subset in the entire data system of the basic user energy usage behavior data and the logical relationship between the subset and other subsets. The basis for grouping includes: performing a first-level grouping according to the source of the basic user energy usage behavior data to distinguish data from different data sources, and then performing a second-level grouping according to the type of the basic user energy usage behavior data within each first-level group, and then performing a third-level grouping according to the time characteristics or energy-consuming device type of the basic user energy usage behavior data within each second-level group; For each subset, according to the intra-group feature derivation rules in the adapted and adjusted feature derivation strategy, the relationship between the internal data of the subset is used to perform basic feature derivation operations to generate the intra-group basic derived feature set of each subset, wherein the intra-group feature derivation rules include basic mathematical operations on numerical data, for a subset with time series data, the difference and ratio between adjacent data points are calculated to reflect the change of the data points over time, and for a subset with category information, basic category-related features are generated according to the relationship between the category information, and the basic category-related features include category proportion and category conversion frequency; According to the inter-group relationship feature derivation rules in the adapted feature derivation strategy, the relationship between different subsets is analyzed to generate inter-group relationship features; According to the multi-feature fusion derivation rules in the adapted and adjusted feature derivation strategy, the derived features of each subset are fused to generate a multi-feature fusion derived feature set of each subset, wherein the derived features include the basic derived features within the group and the relationship features between groups; According to the feature optimization rules in the adapted and adjusted feature derivation strategy, the multi-feature fusion derived feature set of each subset is optimized, and the optimization operation includes feature standardization processing and feature smoothing processing; According to the feature screening rules in the adapted feature derivation strategy, the optimized subset derived feature sets are screened, and after the optimized and screened subset derived feature sets are globally integrated, the globally integrated derived features are renumbered and indexed to generate at least one derived user energy usage behavior data.

9. The method for analyzing user energy usage profile based on AI according to any one of claims 1 to 6, characterized in that: The step of performing an energy usage profile description on the energy usage behavior data of the target user according to the target user energy usage profile analysis model to generate a target energy usage profile description vector comprises: Decomposing the target user's energy consumption behavior data according to predefined rules to generate multiple data blocks, and extracting features from each data block to generate an initial feature set for each data block, wherein the predefined rules are set based on a time window or a data volume, and the initial feature set includes statistical features and semantic features, wherein the statistical features include average value features, maximum value features, minimum value features, and standard deviation features, and the semantic features include behavioral periodicity features and behavioral habit features of the user's energy consumption behavior; Input each data block and the corresponding initial feature set into the initial layer of the target user energy usage profile analysis model, and convert it into an intermediate feature representation after performing a mapping operation according to the predefined neuron connections and weights of the initial layer. The intermediate feature representation is a feature representation converted in the feature direction for describing the energy usage profile label, integrating the relationship between the initial features in the initial feature set; The intermediate feature representation is transmitted and processed between multiple hidden layers in the target user energy usage profile analysis model, wherein, as the intermediate feature representation is propagated between different hidden layers, each hidden layer further fuses and abstracts the intermediate feature representation, and between adjacent hidden layers, the connection weights between neurons perform weighted combination on the input intermediate feature representation to generate a high-level feature representation; In the output layer of the target user energy consumption profile analysis model, a target energy consumption profile feature set related to the energy consumption profile is screened and combined from the high-level feature representation based on the pre-learned energy consumption profile construction mode, and the target energy consumption profile feature set includes a time distribution pattern of energy consumption, a total amount feature of energy consumption, and a stability feature of energy consumption; The initially constructed target energy usage portrait feature set is sorted and encoded to generate a target energy usage portrait description vector, wherein each dimension of the target energy usage portrait description vector corresponds to an energy usage portrait feature in the target energy usage portrait feature set, wherein for numerical energy usage portrait features, their numerical values ​​are directly mapped to the corresponding dimensions of the target energy usage portrait description vector, and for categorical energy usage portrait features, they are converted into vector form using one-hot encoding.

10. An AI-based user energy usage profile analysis system, characterized in that: The AI-based user energy usage portrait analysis system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI-based user energy usage portrait analysis method described in any one of claims 1 to 9 above.

Images