Load temperature decoupling method and system based on user feature analysis
By using a user feature analysis method, a general decoupling model is obtained and the decoupling residual is corrected by combining industry temperature sensitivity and power consumption fluctuation rate. The user groups that are difficult to decouple are identified, and incremental learning is used to obtain a dedicated model. This solves the problem of insufficient decoupling accuracy of the general model in heterogeneous users and achieves high-precision load-temperature decoupling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID SHANGHAI MUNICIPAL ELECTRIC POWER CO
- Filing Date
- 2026-05-14
- Publication Date
- 2026-07-21
AI Technical Summary
In existing technologies, general machine learning models trained based on historical big data have poor load-temperature decoupling accuracy when faced with significant differences between user characteristics and the model training subjects, and cannot effectively separate temperature-sensitive loads from baseline loads.
By acquiring a general decoupling model, preliminary decoupling is performed based on the current temperature and user information. The decoupling residual is corrected by combining industry temperature-sensitive parameters and power capacity fluctuation rate. User groups that are difficult to decouple are identified. A dedicated decoupling sub-model is obtained through incremental learning. The load temperature is decoupled by combining the general and dedicated models.
It significantly improves the accuracy and adaptability of load-temperature decoupling, realizes the transition from generalization to precise adaptation, and enhances the decoupling accuracy for heterogeneous users.
Smart Images
Figure CN122196480B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of load decoupling technology in power distribution networks, specifically to a load-temperature decoupling method and system based on user characteristic analysis. Background Technology
[0002] Power load forecasting and refined analysis are crucial for ensuring stable power grid operation and achieving efficient energy management. Effectively separating temperature-sensitive loads affected by temperature changes from baseline loads unaffected by temperature—a process known as load-temperature decoupling—is fundamental to improving the accuracy of load analysis. Current technologies commonly employ general machine learning or statistical models trained on historical big data to achieve this decoupling. However, in the complex real-world power environment, users exhibit significant heterogeneity in their load patterns due to differences in industry, production scale, energy consumption habits, equipment composition, and geographical location. Different users respond significantly differently to temperature changes. When dealing with users whose characteristics differ greatly from the model's training data, the decoupling residuals of general models increase significantly, resulting in poor decoupling accuracy. Summary of the Invention
[0003] This application provides a load-temperature decoupling method and system based on user feature analysis, which is used to address the technical problem of insufficient accuracy in load-temperature decoupling in the prior art.
[0004] In view of the above problems, this application provides a load-temperature decoupling method and system based on user characteristic analysis.
[0005] In a first aspect, this application provides a load-temperature decoupling method based on user characteristic analysis, the method comprising:
[0006] Obtain a general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results; Based on the preliminary decoupling results, the decoupling residuals are obtained, and the decoupling residuals are corrected based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information to obtain the decoupling effect parameters. Users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult-to-decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult-to-decoupling user group, the best general decoupling sub-model is selected for incremental learning to obtain a dedicated decoupling sub-model. The dedicated decoupling sub-model is used, and combined with general decoupling models other than the dedicated decoupling sub-model, to decouple the load and temperature of difficult-to-decouple user groups and obtain the load and temperature decoupling results.
[0007] Secondly, this application provides a load-temperature decoupling system based on user characteristic analysis, including: The preliminary decoupling module is used to obtain a general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results. The decoupling effect analysis module is used to obtain the decoupling residual based on the preliminary decoupling result, and to correct the decoupling residual based on the industry temperature sensitivity parameter and power capacity fluctuation rate in the current user information, so as to obtain the decoupling effect parameter. The dedicated model acquisition module is used to add users whose decoupling effect parameters are less than the decoupling effect threshold into the difficult-to-decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult-to-decoupling user group, the best general decoupling sub-model is selected for incremental learning to obtain the dedicated decoupling sub-model. The load-temperature decoupling module is used to perform load-temperature decoupling for difficult-to-decouple user groups by employing the dedicated decoupling sub-model and combining it with a general decoupling model other than the dedicated decoupling sub-model, and to obtain the load-temperature decoupling results.
[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes a load-temperature decoupling method and system based on user feature analysis. By dynamically integrating a general decoupling model with user-specific decoupling sub-models, and intelligently identifying and adapting to user groups with difficult decoupling based on user features, it significantly improves the overall accuracy of load-temperature decoupling and the model's adaptability to heterogeneity among different users. Compared with traditional methods, this application combines user feature analysis, model performance evaluation, and personalized incremental learning, achieving an effective transition from generalized to precise adaptation in the field of load-temperature decoupling. While ensuring the model's broad applicability, it significantly improves the accuracy of temperature load decoupling for heterogeneous users. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 This is a flowchart illustrating the load-temperature decoupling method based on user feature analysis provided in an embodiment of this application.
[0011] Figure 2 A schematic diagram of the load-temperature decoupling system based on user feature analysis provided in this application embodiment.
[0012] The components represented by each number in the attached diagram are explained below: The module includes a preliminary decoupling module 100, a decoupling effect analysis module 200, a dedicated model acquisition module 300, and a load-temperature decoupling module 400. Detailed Implementation
[0013] This application provides a load-temperature decoupling method and system based on user feature analysis, which is used to address the technical problem of insufficient accuracy in load-temperature decoupling in the prior art.
[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0015] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.
[0016] Example 1, as Figure 1 As shown, this application provides a load-temperature decoupling method based on user characteristic analysis, wherein the method includes: S10: Obtain the general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results.
[0017] In existing technologies, load-temperature decoupling typically relies on a pre-trained general model, whose goal is to directly output the decoupled load components given the current temperature and the current user. Although the general model covers a large number of samples during training, its averaging characteristics result in poor decoupling capability and accuracy for specific single scenarios.
[0018] Step S10 in the method provided in this application embodiment includes: Obtain a general decoupling model; Among them, obtaining a general decoupling model includes: The historical temperature, user tags, and historical workload of multiple users are obtained to form a training sample set; Sampling with replacement is performed on the training sample set to obtain multiple training sample subsets; Based on machine learning, multiple general decoupling sub-models are constructed; Multiple general decoupling sub-models are obtained by training a subset of training samples and training them separately until they converge. Integrate the general decoupling sub-models to obtain a general decoupling model. The output of the general decoupling model includes the candidate decoupling results of each sub-model and the integrated decoupling result. The integrated decoupling result is the average of the candidate decoupling results. Obtain the current temperature and current user tag from the current user information. Based on the frequency of the current temperature in the training sample set and combined with the power consumption fluctuation rate in the current user information, calculate the sub-model selection coefficient K. Among them, the sub-model selection coefficient K is calculated based on the frequency of the current temperature in the training sample set and combined with the electricity capacity fluctuation rate in the current user information, including: Obtain the frequency of the current temperature in the training sample set; The analysis difficulty parameter is obtained by subtracting the frequency from 1. Obtain the historical electricity consumption capacity sequence stored in the current user information, calculate the ratio of the standard deviation to the mean of the historical electricity consumption capacity sequence, and obtain the electricity consumption capacity volatility. The analysis difficulty parameter and the power capacity fluctuation rate are weighted and summed, and then multiplied by the basic selection coefficient to obtain the sub-model selection coefficient K.
[0019] Input the current temperature and current user tag into the general decoupling model, call K general decoupling sub-models to decouple the load and temperature, and obtain preliminary decoupling results. The preliminary decoupling results include K alternative decoupling results and integrated decoupling results.
[0020] In this embodiment of the application, a general decoupling model is obtained, and load-temperature decoupling is performed based on the current temperature and current user information to obtain preliminary decoupling results.
[0021] Specifically, obtain a general decoupling model.
[0022] First, historical temperatures, user tags, and historical loads of multiple users are acquired to form a training sample set. These users include past data from power grid users such as chemical plants, commercial buildings, and residential communities. Daily temperature records are labeled as historical temperatures, user industry categories, and corresponding power load readings as historical loads. For example, a sliding window of length 3 is used. For each sampling time, the load values at that time and one time before and after it are taken, and the median of these three load values is calculated as the filtered load value for that time. This median filtering process effectively removes isolated noise points from the load data while preserving the overall trend of load changes, resulting in a preprocessed training sample set.
[0023] Furthermore, sampling with replacement is performed on the training sample set to obtain multiple training sample subsets. Specifically, sampling with replacement is performed on the above training sample set, that is, each time a data record is randomly selected and then put back into the set, so that it may be selected again. This process is repeated multiple times to obtain several training sample subsets with overlapping but not completely identical content.
[0024] Furthermore, a subset of training samples is used to train each general decoupling sub-model until convergence, resulting in multiple general decoupling sub-models. Specifically, a decision tree model is used as the sub-model. Each decision tree model consists of a root node, multiple intermediate nodes, and multiple leaf nodes. The maximum depth of the tree is set to 5, and the minimum number of samples required for each intermediate node to split is 10. The model input consists of numerical features of temperature and user labels, and the model output is a specific decoupling load prediction value. Multiple subsets of training samples are used to train each general decoupling sub-model separately until convergence, resulting in multiple general decoupling sub-models.
[0025] Furthermore, the general decoupling sub-models are integrated to obtain a general decoupling model. The output of the general decoupling model includes the candidate decoupling results of each sub-model and the integrated decoupling result, which is the mean of the candidate decoupling results. These trained general decoupling sub-models are integrated to construct the final general decoupling model. When input is received, the general decoupling sub-models are invoked to perform calculations. The output contains two parts: the candidate decoupling results are the temperature-sensitive load values and non-temperature-sensitive load values calculated by each of the general decoupling sub-models; the integrated decoupling result is the arithmetic mean of the candidate decoupling results.
[0026] Furthermore, the current temperature and current user tag are obtained from the current user information. Based on the frequency of the current temperature in the training sample set and combined with the electricity capacity fluctuation rate in the current user information, the sub-model selection coefficient K is calculated, including: Obtain the frequency of the current temperature in the training sample set; The analysis difficulty parameter is obtained by subtracting the frequency from 1. Obtain the historical electricity consumption capacity sequence stored in the current user information, calculate the ratio of the standard deviation to the mean of the historical electricity consumption capacity sequence, and obtain the electricity consumption capacity volatility. The analysis difficulty parameter and the power capacity fluctuation rate are weighted and summed, and then multiplied by the basic selection coefficient to obtain the sub-model selection coefficient K.
[0027] Among them, the historical electricity consumption sequence refers to the electricity consumption sequence of the current user within a preset time period in the past.
[0028] Specifically, when decoupling from a current user, the current temperature value and the user's tag are first obtained. Further, based on the frequency of the current temperature in the training sample set and combined with the electricity capacity fluctuation rate in the current user information, the sub-model selection coefficient K is calculated.
[0029] First, obtain the frequency of the current temperature in the training sample set. Subtract this frequency from 1 to obtain the analysis difficulty parameter. For example, calculate the frequency of the current temperature value (e.g., 30 degrees) in all previously collected historical temperature data. For instance, the frequency of 30 degrees in history is 0.05, or 5%. Subtracting this frequency from 1 gives the analysis difficulty parameter = 1 - 0.05 = 0.95.
[0030] Furthermore, the historical electricity consumption sequence stored in the current user information is obtained, and the ratio of the standard deviation to the mean of the historical electricity consumption sequence is calculated to obtain the electricity consumption volatility. For example, if the average electricity consumption of a chemical plant over the past 90 days is 2000 kVA and the standard deviation is 300 kVA, then the electricity consumption volatility = 300 / 2000 = 0.15.
[0031] Furthermore, the analysis difficulty parameter and the power capacity volatility are weighted and summed, then multiplied by a base selection coefficient to obtain the sub-model selection coefficient K. For example, the analysis difficulty parameter (0.95) and the power capacity volatility parameter (0.15) are weighted and summed. For instance, if the current user's power capacity volatility is relatively low, the weight of the analysis difficulty parameter needs to be emphasized. The weight of the analysis difficulty parameter can be set to 0.7, and the weight of the power capacity volatility parameter to 0.3. The weighted sum is 0.95 × 0.7 + 0.15 × 0.3 = 0.665 + 0.045 = 0.71. Multiplying this by a preset base selection coefficient (e.g., 30) yields K = 30 × 0.71 = 21.3. If the result is not an integer, the decimal is discarded, resulting in K = 21. The more uncommon the current temperature or the more unstable the user's power consumption, the larger the sub-model selection coefficient K becomes, requiring more sub-models to be called, thus enhancing the robustness of decoupling.
[0032] Furthermore, the current temperature and current user tag are input into the general decoupling model, and K general decoupling sub-models are invoked to perform load-temperature decoupling, obtaining preliminary decoupling results. These preliminary decoupling results include K alternative decoupling results and an integrated decoupling result. For example, the current temperature and current user tag are input into the general decoupling model. Based on the value of K, the model invokes K general decoupling sub-models to perform decoupling calculations. These K sub-models will output K temperature-sensitive load prediction values; for example, if K=28, 28 alternative decoupling results are obtained. Simultaneously, the model calculates the average of these 28 values to obtain an integrated decoupling result. These 28 alternative decoupling results and 1 integrated decoupling result together constitute the preliminary decoupling result for this user at this temperature.
[0033] The method provided in this application does not simply decouple based on the current temperature and the current user. Instead, it dynamically determines a sub-model selection coefficient K based on the characteristics of the current temperature in historical data, and then calls K of the most relevant general decoupling sub-models for parallel computation. By introducing the diverse outputs of multiple sub-models, not only is the random bias that may exist in a single model smoothed out through integration, improving the stability of the initial results, but more importantly, the multiple alternative results themselves constitute a solution space of decoupling possibilities. This provides a crucial comparative basis for subsequent steps to analyze decoupling consistency and evaluate the model's performance in this user scenario.
[0034] S20: Based on the preliminary decoupling results, obtain the decoupling residuals, and based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information, correct the decoupling residuals to obtain the decoupling effect parameters.
[0035] After obtaining preliminary decoupling results, existing technologies often lack an effective and quantitative standard to judge whether the decoupling is successful and accurate for the current user, and are even less able to distinguish the differences in decoupling difficulty between different user groups.
[0036] Step S20 in the method provided in this application embodiment includes: Based on current user information, obtain the current load; Based on the deviation of the integrated decoupling results from the current load and the preliminary decoupling results, obtain the decoupling deviation parameter; Obtain user characteristics from the current user information, including industry category and electricity capacity fluctuation rate; Based on the industry category in user characteristics, analyze and obtain industry temperature sensitivity parameters; The decoupling deviation parameters are corrected using industry temperature sensitivity parameters and power capacity fluctuation rate to obtain decoupling effect parameters.
[0037] In this embodiment of the application, based on the preliminary decoupling results, the decoupling residual is obtained, and based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information, the decoupling residual is corrected to obtain the decoupling effect parameters.
[0038] Specifically, firstly, based on current user information, the current load is obtained. For example, at 2 PM, the current load of a chemical plant is 2000 kVA.
[0039] Furthermore, a decoupling deviation parameter is obtained based on the deviation between the current load and the integrated decoupling result in the preliminary decoupling results. For example, the decoupling deviation parameter = |current load - integrated decoupling result| / current load, where non-temperature-sensitive loads are obtained based on the minimum load of the current user. The closer the decoupling deviation parameter is to 0, the better the preliminary decoupling result matches the actual result; the closer it is to 1, the greater the deviation.
[0040] Furthermore, user characteristics are obtained from the current user information, including industry category and electricity capacity fluctuation rate. For example, the user's industry category, such as chemical, textile, or commercial, is obtained, and the electricity capacity, such as the rated capacity of the user's transformer, is obtained.
[0041] Furthermore, based on the industry category in user characteristics, industry-specific temperature sensitivity parameters are obtained through analysis. For example, based on a sample of users belonging to the same industry category, such as "chemical industry," in historical data, the strength of the correlation between the overall power load and temperature changes in that industry can be analyzed. This can be achieved by calculating the absolute value of the correlation coefficient between historical load data and historical temperature data for that industry, and then normalizing this absolute value to a range of 0 to 1 to obtain the industry's temperature sensitivity parameter. For instance, for the chemical industry, the production process may be mainly controlled by technology and not sensitive to temperature changes, so the calculated temperature sensitivity parameter might be low, at 0.2; for commercial buildings, air conditioning load accounts for a large proportion and is sensitive to temperature, so its temperature sensitivity parameter would be higher, at 0.7.
[0042] Furthermore, the decoupling deviation parameter is corrected using industry temperature sensitivity parameters and electricity capacity volatility to obtain the decoupling effect parameter. The decoupling effect parameter = decoupling deviation parameter × [1 + ln(1 + industry temperature sensitivity parameter) × (1 + electricity capacity volatility)], where ln is the natural logarithm function, and electricity capacity volatility is used to characterize the stability of the user's load. For example, if a user's decoupling deviation parameter is 0.2, the industry temperature sensitivity parameter is 0.7, and the electricity capacity volatility is 0.15, then ln(1 + 0.7) = ln1.7 ≈ 0.531, 1 + ln(1 + industry temperature sensitivity parameter) × (1 + electricity capacity volatility) = 1 + 0.531 × (1 + 0.15) = 1 + 0.531 × 1.15 = 1 + 0.61065 = 1.61065, and the final decoupling effect parameter = 0.2 × 1.61065 ≈ 0.322. The nonlinear correction formula makes the influence of the temperature-sensitive parameter grow logarithmically, and the growth slows down when the temperature sensitivity is high, which is more in line with the physical law that the influence of temperature has a saturation effect. At the same time, it is combined with the user load stability for comprehensive correction, so as to more accurately reflect the actual decoupling difficulty of different users in complex scenarios.
[0043] By analyzing the general temperature sensitivity patterns of specific industries, industry-specific temperature sensitivity parameters are obtained, and these parameters are used to correct or standardize the original decoupling bias. The obtained decoupling effect parameters are no longer isolated, absolute numbers, but rather relative evaluation indicators that combine model performance with the user's own baseline decoupling difficulty. This provides an accurate and reliable basis for the next step of automatically identifying and classifying user groups.
[0044] S30: Add users whose decoupling effect parameter is less than the decoupling effect threshold to the difficult-to-decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult-to-decoupling user group, select the best general decoupling sub-model for incremental learning to obtain a dedicated decoupling sub-model.
[0045] Once users with poor decoupling performance are identified, continuing to use the original general model will not solve the accuracy problem; training a completely new dedicated model from scratch for each such user will require a lot of computing resources, time, and labeled data, which is inefficient.
[0046] Step S30 in the method provided in this application embodiment includes: Users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult decoupling user group, which includes multiple difficult decoupling users. The determination of the decoupling effect threshold includes: Obtain multiple decoupling effect parameters for each user; Cluster analysis of decoupling effect parameters was performed on each user to form multiple user clusters; Based on multiple decoupling effect parameters of user clusters, obtain multiple decoupling effect thresholds for each user cluster; Based on user characteristics in the current user information, similar user clusters are obtained, and the decoupling effect threshold of similar user clusters is used as the decoupling effect threshold. Obtain user characteristics, historical temperature, and historical load of users who are difficult to decouple, and integrate them into an incremental dataset; Obtain the power capacity volatility of users who are difficult to decouple, and determine the learning rate for incremental learning. The larger the power capacity volatility, the more inversely proportional the learning rate. Select the best general decoupling sub-model and use an incremental dataset to incrementally learn the best general decoupling sub-model to obtain a custom decoupling sub-model; The acquisition of the optimal general decoupling sub-model includes: Obtain K alternative decoupling results from the preliminary decoupling results; Calculate the current deviation between the K alternative decoupling results and the current load, and calculate the average decoupling deviation of the general decoupling sub-model corresponding to each alternative decoupling result for similar users in historical decoupling. The current deviation is weighted and fused with the average decoupling deviation to obtain the comprehensive decoupling deviation; The candidate decoupling result with the smallest deviation after weighted fusion is selected as the optimal candidate decoupling result; The general decoupling sub-model corresponding to the best alternative decoupling result is taken as the best general decoupling sub-model.
[0047] In this embodiment, users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult-to-decoupling user group. Combining user characteristics, historical temperature and historical load in the current user information, the best general decoupling sub-model is selected for incremental learning to obtain a dedicated decoupling sub-model.
[0048] Specifically, users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult-to-decoupling user group, which includes multiple difficult-to-decoupling users.
[0049] First, the decoupling effect threshold is obtained.
[0050] Obtain multiple decoupling effect parameters for each user. For example, obtain multiple decoupling effect parameters for multiple users with different user characteristics.
[0051] Furthermore, decoupling effect parameter clustering analysis is performed on each user to form multiple user clusters. For example, the K-means algorithm is used to cluster these users based on their decoupling effect parameters. Through clustering, users with similar decoupling effect parameters are grouped into the same group, forming multiple user clusters. Specifically, the decoupling effect parameter of each user is used as a clustering feature, combined with industry category coding and electricity capacity values from the user features for multi-dimensional feature clustering. The number of clusters is set to 3 to 5, and the optimal number of clusters is determined using the elbow rule, i.e., calculating the sum of squared intra-cluster errors under different cluster numbers, and selecting the cluster number corresponding to the inflection point of the curve as the final number of clusters. For example, the elbow rule determines that the optimal number of clusters is 4, that is, dividing 1000 users into 4 user clusters, with users in each cluster having similar decoupling effect parameters and similar electricity consumption characteristics. For each user cluster, the average decoupling effect parameter of all users in that cluster is calculated, and this average is used as the decoupling effect threshold for that user cluster.
[0052] Furthermore, based on user characteristics in the current user information, similar user clusters are obtained, and the decoupling effect threshold of the similar user clusters is used as the decoupling effect threshold. When it is necessary to determine the decoupling effect for a specific current user, such as a chemical plant, the user cluster most similar to its characteristics is retrieved based on the user's characteristics, and the decoupling effect threshold of that cluster is used as the decoupling effect threshold for the current user. For example, a chemical plant with low temperature sensitivity may be classified into a user cluster with "generally small decoupling effect parameters," and the threshold for this cluster may be a small value.
[0053] Furthermore, users whose decoupling effect parameters are less than the decoupling effect threshold are added to the difficult-to-decoupling user group, which includes multiple users. For example, the decoupling effect parameter of the current user is compared with the obtained decoupling effect threshold. If the parameter value is less than the threshold of its user cluster, the user is added to the difficult-to-decoupling user group. That is, for the type to which this user belongs, its current decoupling accuracy has not reached the group average level and requires special handling.
[0054] Furthermore, user characteristics, historical temperature, and historical load of users with difficulty in decoupling are obtained and integrated into an incremental dataset. Historical temperature and load data of users with difficulty in decoupling over a period of time, such as the past three months, are extracted and integrated with their user characteristics to form an incremental dataset specifically for model relearning.
[0055] Furthermore, the power capacity volatility of users with unstable power consumption is obtained to determine the learning rate for incremental learning. Specifically, the learning rate is inversely proportional to the power capacity volatility. The learning rate is determined based on the power capacity volatility of these users, with a larger volatility resulting in a smaller learning rate to control the magnitude of model updates. For example, if the baseline learning rate is 0.1 and the power capacity volatility is 0.15, then the learning rate = baseline learning rate × (1 - power capacity volatility) = 0.1 × 0.85 = 0.085. If another user's power capacity volatility is 0.3, then the learning rate = 0.1 × 0.7 = 0.07. Users with more unstable power consumption require a smaller learning rate to avoid the model being excessively affected by noisy data, achieving more robust personalized adaptation.
[0056] Furthermore, incremental learning is performed on the best general decoupling sub-model using an incremental dataset to obtain a custom decoupling sub-model.
[0057] The acquisition of the optimal general decoupling sub-model includes: First, obtain K alternative decoupling results from the preliminary decoupling results.
[0058] Furthermore, the current deviations between the K alternative decoupling results and the current load are calculated, and the average decoupling deviation of the general decoupling sub-model corresponding to each alternative decoupling result for similar users in historical decoupling is calculated. For example, the absolute differences between the K alternative decoupling results and the current load are calculated to obtain the current deviation value. For instance, the current deviation of sub-model A is 20 kVA, the current deviation of sub-model B is 20 kVA, and the current deviation of sub-model C is 40 kVA.
[0059] Furthermore, the current deviation and the average decoupling deviation are weighted and fused to obtain the comprehensive decoupling deviation. Specifically, the historical deviation sequence of the general decoupling sub-model for decoupling the current user's similar user cluster is obtained, and its average value is calculated as the historical average deviation. Here, users in the similar user cluster are similar users to the current user. For example, the historical average deviation of sub-model A for similar users is 15kVA, the historical average deviation of sub-model B is 25kVA, and the historical average deviation of sub-model C is 10kVA.
[0060] The candidate decoupling result with the smallest bias after weighted fusion is selected as the optimal candidate decoupling result. For example, regarding the selection of sub-models, the current user's current bias is more important. Therefore, the current bias weight is set to 0.6, and the historical average bias weight is set to 0.4. Then, the comprehensive bias of sub-model A is 20 × 0.6 + 15 × 0.4 = 18 kVA; the comprehensive bias of sub-model B is 20 × 0.6 + 25 × 0.4 = 22 kVA; and the comprehensive bias of sub-model C is 40 × 0.6 + 10 × 0.4 = 28 kVA. Comparison shows that sub-model A has the smallest comprehensive bias; therefore, the candidate decoupling result corresponding to sub-model A is selected as the optimal candidate decoupling result. Although the current bias of sub-model A and sub-model B is both 20 kVA, sub-model A has more stable historical performance and is considered the best choice after comprehensive evaluation.
[0061] Furthermore, the general decoupling sub-model corresponding to the optimal candidate decoupling result is taken as the best general decoupling sub-model.
[0062] Furthermore, incremental learning is performed on the optimal general decoupling sub-model using an incremental dataset to obtain a personalized decoupling sub-model. Specifically, while keeping the original decision tree structure unchanged, the predicted values stored in each leaf node of the tree are updated using the incremental dataset. The update method is exemplarily set to calculate the average load of all new and old data corresponding to that leaf node. Simultaneously, based on the incremental dataset, and provided the original stopping condition is met, the nodes are locally split, growing new branches. After completing this incremental learning process, the optimal general decoupling sub-model is updated into a new decision tree that incorporates the user's personalized data; this new decision tree is the personalized decoupling sub-model for this difficult-to-decouple user.
[0063] By introducing incremental learning and optimal sub-model selection strategies, this approach efficiently achieves personalized model customization for users with difficult decoupling needs, balancing accuracy improvement with resource conservation. First, instead of blindly training a new model for each user, the method selects the best-performing general decoupling sub-model from a proven library of general decoupling models, using it as the foundation. This optimal general decoupling sub-model already possesses strong learning capabilities, and its output shows minimal deviation from the actual load in the initial decoupling, indicating that its initial state is relatively close to the user's potential pattern. Furthermore, incremental learning is performed on the optimal general decoupling sub-model using an incremental dataset comprised of the user's specific characteristics, historical temperature, and historical load. The resulting personalized decoupling sub-model inherits the universal knowledge framework of the general model while deeply integrating the user's personalized load-temperature response characteristics. This achieves an improvement from general approximation to personalized adaptation with relatively low additional computational cost, avoiding the computational waste associated with full retraining.
[0064] S40: Use a dedicated decoupling sub-model and combine it with a general decoupling model other than the dedicated decoupling sub-model to decouple the load temperature of difficult-to-decouple user groups and obtain the load temperature decoupling results.
[0065] Having developed a dedicated decoupling sub-model for users with difficult decoupling needs, completely abandoning the general model and using only the dedicated model may lead to overfitting or insufficient stability on small samples because the dedicated model is trained solely on historical data from a single user. Conversely, if the results of the general model are still the primary focus, the improvement effect of the dedicated model cannot be fully realized.
[0066] Step S40 in the method provided in this application embodiment includes: Based on the decoupling effect parameters, obtain the sub-model selection coefficient Q; Q general decoupling sub-models, including the exclusive decoupling sub-model and the general decoupling model other than the exclusive decoupling sub-model, are selected to decouple the load temperature of the difficult-to-decouple user group, and the exclusive decoupling results and general decoupling results are obtained. The general decoupling result is the average of the candidate decoupling results of the Q general decoupling sub-models. The load-temperature decoupling result is obtained by weighting the specific decoupling result and the general decoupling result. The weights in the weighting calculation are obtained based on industry temperature sensitivity parameters.
[0067] In this embodiment, a dedicated decoupling sub-model is used in combination with a general decoupling model other than the dedicated decoupling sub-model to decouple the load temperature of difficult-to-decouple user groups and obtain the load temperature decoupling result.
[0068] Specifically, firstly, based on the decoupling effect parameter, the sub-model selection coefficient Q is obtained. A larger decoupling effect parameter indicates a better initial decoupling effect, thus reducing the number of general sub-models called to improve computational efficiency; a smaller decoupling effect parameter indicates a less than ideal initial decoupling effect, thus increasing the number of general sub-models called to gather more model intelligence to improve accuracy. For example, Q = (1 - decoupling effect parameter) × basic selection coefficient. The basic selection coefficient is a preset constant. For instance, if a user's decoupling effect parameter is 0.15 (a small value indicating poor decoupling), and the preset basic selection coefficient is 10, then Q = (1 - 0.15) × 10 = 8.5. Rounding down, we get Q = 9, meaning 9 general sub-models are called in the final calculation to improve accuracy.
[0069] Furthermore, Q general decoupling sub-models from the dedicated decoupling sub-model and other general decoupling sub-models are selected to decouple the load temperature of the difficult-to-decouple user group, obtaining dedicated decoupling results and general decoupling results. The general decoupling result is the average of the Q candidate decoupling results from the general decoupling sub-models. A dedicated decoupling sub-model generated for this user is selected, and simultaneously, Q sub-models are randomly selected from all general decoupling sub-models. The current temperature and the current user label are simultaneously input into both the dedicated decoupling sub-model and the selected Q general decoupling sub-models. The dedicated decoupling sub-model calculates based on its personalized rules and outputs a temperature-sensitive load prediction value, which is called the dedicated decoupling result. Simultaneously, each of the Q general decoupling sub-models outputs a temperature-sensitive load prediction value, resulting in Q candidate decoupling results. The arithmetic mean of these Q candidate decoupling results is then calculated to obtain the general decoupling result.
[0070] Furthermore, the specific decoupling results and the general decoupling results are weighted and calculated to obtain the load-temperature decoupling result. The weights in the weighting calculation are based on industry-specific temperature sensitivity parameters. For example, the load-temperature decoupling result = (specific decoupling result × industry-specific temperature sensitivity parameter) + [general decoupling result × (1 - industry-specific temperature sensitivity parameter)]. For instance, if a user is a chemical plant with an industry-specific temperature sensitivity parameter of 0.2, then the weight of its specific decoupling result is 0.2, and the weight of its general decoupling result is 0.8. Through this weighted calculation, the final output load-temperature decoupling result includes both user-specific information learned individually and the robustness of the general model, with the fusion ratio adaptively adjusted according to the inherent temperature sensitivity of the industry.
[0071] In this embodiment, an adaptive weighted fusion strategy is employed to achieve complementary advantages between the general decoupling model and the specific decoupling sub-model, thereby obtaining the optimal load-temperature decoupling result. The obtained load-temperature decoupling result is not a simple average, but rather a weighted calculation of the identification results, with the weights scientifically allocated based on industry-specific temperature sensitivity parameters that represent the essential characteristics of the user. This fusion mechanism ensures that users with highly temperature-sensitive industry applications can assign higher weights to their specific model results to capture subtle changes; while for users with low temperature sensitivity, the robust output of the general model is relied upon more. Ultimately, this fusion output overcomes the accuracy limitations of a single general model for users with difficult decoupling needs, while mitigating the overfitting risk that may exist with a single specific model, comprehensively improving the accuracy of the load-temperature decoupling result.
[0072] Example 2, as Figure 2 As shown, based on the same inventive concept as the load-temperature decoupling method based on user feature analysis provided in Embodiment 1, this embodiment of the invention also provides a load-temperature decoupling system based on user feature analysis, including: The preliminary decoupling module 100 is used to obtain a general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results. The decoupling effect analysis module 200 is used to obtain the decoupling residual based on the preliminary decoupling results, and to correct the decoupling residual based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information, so as to obtain the decoupling effect parameters. The dedicated model acquisition module 300 is used to add users whose decoupling effect parameters are less than the decoupling effect threshold into the difficult decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult decoupling user group, the best general decoupling sub-model is selected for incremental learning to obtain the dedicated decoupling sub-model. The load-temperature decoupling module 400 is used to decouple the load and temperature of user groups that are difficult to decouple by employing a dedicated decoupling sub-model and combining it with a general decoupling model other than the dedicated decoupling sub-model, and to obtain the load-temperature decoupling results.
[0073] In one embodiment, the preliminary decoupling module 100 is further configured to: Obtain a general decoupling model; Among them, obtaining a general decoupling model includes: The historical temperature, user tags, and historical workload of multiple users are obtained to form a training sample set; Sampling with replacement is performed on the training sample set to obtain multiple training sample subsets; Based on machine learning, multiple general decoupling sub-models are constructed; Multiple general decoupling sub-models are obtained by training a subset of training samples and training them separately until they converge. Integrate the general decoupling sub-models to obtain a general decoupling model. The output of the general decoupling model includes the candidate decoupling results of each sub-model and the integrated decoupling result. The integrated decoupling result is the average of the candidate decoupling results. Obtain the current temperature and current user tag from the current user information. Based on the frequency of the current temperature in the training sample set and combined with the power consumption fluctuation rate in the current user information, calculate the sub-model selection coefficient K. Among them, the sub-model selection coefficient K is calculated based on the frequency of the current temperature in the training sample set and combined with the electricity capacity fluctuation rate in the current user information, including: Obtain the frequency of the current temperature in the training sample set; The analysis difficulty parameter is obtained by subtracting the frequency from 1. Obtain the historical electricity consumption capacity sequence stored in the current user information, calculate the ratio of the standard deviation to the mean of the historical electricity consumption capacity sequence, and obtain the electricity consumption capacity volatility. The analysis difficulty parameter and the power capacity fluctuation rate are weighted and summed, and then multiplied by the basic selection coefficient to obtain the sub-model selection coefficient K.
[0074] Input the current temperature and current user tag into the general decoupling model, call K general decoupling sub-models to decouple the load and temperature, and obtain preliminary decoupling results. The preliminary decoupling results include K alternative decoupling results and integrated decoupling results.
[0075] In one embodiment, the decoupling effect analysis module 200 is further configured to: Based on current user information, obtain the current load; Based on the deviation of the integrated decoupling results from the current load and the preliminary decoupling results, obtain the decoupling deviation parameter; Obtain user characteristics from the current user information, including industry category and electricity capacity fluctuation rate; Based on the industry category in user characteristics, analyze and obtain industry temperature sensitivity parameters; The decoupling deviation parameters are corrected using industry temperature sensitivity parameters and power capacity fluctuation rate to obtain decoupling effect parameters.
[0076] In one embodiment, the dedicated model acquisition module 300 is further used for: Users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult decoupling user group, which includes multiple difficult decoupling users. The determination of the decoupling effect threshold includes: Obtain multiple decoupling effect parameters for each user; Cluster analysis of decoupling effect parameters was performed on each user to form multiple user clusters; Based on multiple decoupling effect parameters of user clusters, obtain multiple decoupling effect thresholds for each user cluster; Based on user characteristics in the current user information, similar user clusters are obtained, and the decoupling effect threshold of similar user clusters is used as the decoupling effect threshold. Obtain user characteristics, historical temperature, and historical load of users who are difficult to decouple, and integrate them into an incremental dataset; Obtain the power capacity volatility of users who are difficult to decouple, and determine the learning rate for incremental learning. The larger the power capacity volatility, the more inversely proportional the learning rate. Select the best general decoupling sub-model and use an incremental dataset to incrementally learn the best general decoupling sub-model to obtain a custom decoupling sub-model; The acquisition of the optimal general decoupling sub-model includes: Obtain K alternative decoupling results from the preliminary decoupling results; Calculate the current deviation between the K alternative decoupling results and the current load, and calculate the average decoupling deviation of the general decoupling sub-model corresponding to each alternative decoupling result for similar users in historical decoupling. The current deviation is weighted and fused with the average decoupling deviation to obtain the comprehensive decoupling deviation; The candidate decoupling result with the smallest deviation after weighted fusion is selected as the optimal candidate decoupling result; The general decoupling sub-model corresponding to the best alternative decoupling result is taken as the best general decoupling sub-model.
[0077] In one embodiment, the load temperature decoupling module 400 is further configured to: Based on the decoupling effect parameters, obtain the sub-model selection coefficient Q; Q general decoupling sub-models, including the exclusive decoupling sub-model and the general decoupling model other than the exclusive decoupling sub-model, are selected to decouple the load temperature of the difficult-to-decouple user group, and the exclusive decoupling results and general decoupling results are obtained. The general decoupling result is the average of the candidate decoupling results of the Q general decoupling sub-models. The load-temperature decoupling result is obtained by weighting the specific decoupling result and the general decoupling result. The weights in the weighting calculation are obtained based on industry temperature sensitivity parameters.
[0078] In summary, the embodiments of this application have at least the following technical effects: This application proposes a load-temperature decoupling method and system based on user feature analysis. By dynamically integrating a general decoupling model with a user-specific decoupling sub-model, and intelligently identifying and adapting to user groups that are difficult to decouple based on user features, the overall accuracy of load-temperature decoupling and the model's adaptability to the heterogeneity of different users are significantly improved. Specifically, this application introduces user feature analysis and incremental learning mechanisms. First, it uses a general decoupling model to perform preliminary load-temperature decoupling on current user information. Then, based on the decoupling residuals and user characteristics, it calculates the decoupling effect parameters, thereby scientifically identifying user groups that are difficult to decouple due to user heterogeneity. For these groups, this application does not adopt the high-computational-cost method of global model reconstruction. Instead, it selects the best sub-model from the general model and performs incremental learning based on the user's historical temperature and load data to quickly generate a lightweight, dedicated decoupling sub-model. This process fully utilizes existing model knowledge and achieves personalized adaptation to the temperature-sensitive patterns of specific users with low computational overhead. In the final decoupling stage, by organically integrating and weighting the dedicated sub-model with multiple sub-models in the general model, the decoupling results retain the universality advantage of the general model for common users while strengthening the capture of temperature-sensitive characteristics for users who are difficult to decouple, thereby improving the overall reliability and stability of the decoupling results. Compared with traditional methods, this application combines user feature analysis, model performance evaluation and personalized incremental learning to achieve an effective transition from generalized generalization to precise adaptation in the field of load-temperature decoupling. While ensuring the wide applicability of the model, it significantly improves the accuracy of temperature load decoupling for heterogeneous users.
[0079] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0080] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0081] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.
Claims
1. A load-temperature decoupling method based on user characteristic analysis, characterized in that, include: Obtain a general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results; Based on the preliminary decoupling results, the decoupling residuals are obtained, and the decoupling residuals are corrected based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information to obtain the decoupling effect parameters. Users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult-to-decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult-to-decoupling user group, the best general decoupling sub-model is selected for incremental learning to obtain a dedicated decoupling sub-model. The dedicated decoupling sub-model is used, and combined with general decoupling models other than the dedicated decoupling sub-model, to decouple the load and temperature of difficult-to-decouple user groups and obtain the load and temperature decoupling results.
2. The load-temperature decoupling method based on user feature analysis according to claim 1, characterized in that, The process of obtaining a general decoupling model involves decoupling the load by temperature based on the current temperature and current user information, and obtaining preliminary decoupling results, including: Obtain a general decoupling model; Obtain the current temperature and current user tag from the current user information. Based on the frequency of the current temperature in the training sample set and combined with the power consumption fluctuation rate in the current user information, calculate the sub-model selection coefficient K. The current temperature and current user tag are input into the general decoupling model, and K general decoupling sub-models are called to decouple the load and temperature to obtain preliminary decoupling results. The preliminary decoupling results include K alternative decoupling results and integrated decoupling results.
3. The load-temperature decoupling method based on user feature analysis according to claim 2, characterized in that, The process of obtaining the general decoupling model includes: The historical temperature, user tags, and historical workload of multiple users are obtained to form a training sample set; Sampling with replacement is performed on the training sample set to obtain multiple training sample subsets; Based on machine learning, multiple general decoupling sub-models are constructed; Each of the general decoupling sub-models is trained using the aforementioned subset of training samples until convergence, thereby obtaining multiple general decoupling models; The general decoupling sub-models are integrated to obtain a general decoupling model. The output of the general decoupling model includes the candidate decoupling results of each decoupling sub-model and the integrated decoupling result. The integrated decoupling result is the average of the candidate decoupling results.
4. The load-temperature decoupling method based on user feature analysis according to claim 2, characterized in that, The sub-model selection coefficient K is calculated based on the frequency of the current temperature in the training sample set and combined with the electricity capacity fluctuation rate in the current user information, including: Obtain the frequency of the current temperature in the training sample set; The analysis difficulty parameter is obtained by subtracting the frequency from 1. Obtain the historical electricity consumption capacity sequence stored in the current user information, calculate the ratio of the standard deviation to the mean of the historical electricity consumption capacity sequence, and obtain the electricity consumption capacity volatility. The analysis difficulty parameter and the power capacity fluctuation rate are weighted and summed, and then multiplied by the basic selection coefficient to obtain the sub-model selection coefficient K.
5. The load-temperature decoupling method based on user feature analysis according to claim 1, characterized in that, Based on the preliminary decoupling results, a decoupling residual is obtained. Then, based on the industry temperature sensitivity parameters and power capacity fluctuation rate in the current user information, the decoupling residual is corrected to obtain decoupling effect parameters, including: Based on the current user information, obtain the current load; Based on the deviation between the current load and the integrated decoupling result in the preliminary decoupling result, the decoupling deviation parameter is obtained; Obtain user characteristics from the current user information, wherein the user characteristics include industry category and electricity capacity fluctuation rate; Based on the industry category in the user characteristics, industry temperature sensitivity parameters are analyzed and obtained. The decoupling deviation parameter is corrected using the industry temperature sensitivity parameter and the power capacity fluctuation rate to obtain the decoupling effect parameter.
6. The load-temperature decoupling method based on user feature analysis according to claim 1, characterized in that, The process involves adding users whose decoupling effect parameter is less than the decoupling effect threshold to the difficult-to-decoupling user group. Based on the user characteristics, historical temperature, and historical load of the difficult-to-decoupling user group, the optimal general decoupling sub-model is selected for incremental learning to obtain a specific decoupling sub-model, including: Users whose decoupling effect parameter is less than the decoupling effect threshold are added to the difficult-to-decoupling user group, wherein the difficult-to-decoupling user group includes multiple difficult-to-decoupling users; Obtain the user characteristics, historical temperature, and historical load of the users who are difficult to decouple, and integrate them into an incremental dataset; Obtain the power capacity fluctuation rate of the difficult-to-decouple user, and determine the learning rate for incremental learning, wherein the power capacity fluctuation rate and the learning rate are inversely proportional; Select the best general decoupling sub-model, and use the incremental dataset to incrementally learn the best general decoupling sub-model to obtain the exclusive decoupling sub-model.
7. The load-temperature decoupling method based on user feature analysis according to claim 5, characterized in that, Obtaining the decoupling effect threshold includes: Obtain multiple decoupling effect parameters for each user; Cluster analysis of decoupling effect parameters was performed on each user to form multiple user clusters; Based on multiple decoupling effect parameters of the user cluster, multiple decoupling effect thresholds for each user cluster are obtained; Based on user characteristics in the current user information, similar user clusters are obtained, and the decoupling effect threshold of the similar user clusters is used as the decoupling effect threshold.
8. The load-temperature decoupling method based on user feature analysis according to claim 5, characterized in that, The acquisition of the optimal general decoupling sub-model includes: Obtain K alternative decoupling results from the preliminary decoupling results; Calculate the current deviation between the K candidate decoupling results and the current load, and calculate the average decoupling deviation of the general decoupling sub-model corresponding to each candidate decoupling result for similar users in historical decoupling. The current deviation is weighted and fused with the average decoupling deviation to obtain the comprehensive decoupling deviation; The candidate decoupling result with the smallest deviation after weighted fusion is selected as the optimal candidate decoupling result; The general decoupling sub-model corresponding to the optimal candidate decoupling result is taken as the best general decoupling sub-model.
9. The load-temperature decoupling method based on user feature analysis according to claim 1, characterized in that, The process of using the dedicated decoupling sub-model and combining it with a general decoupling model other than the dedicated decoupling sub-model to decouple the load and temperature of difficult-to-decouple user groups, and obtaining the load and temperature decoupling results, includes: Based on the decoupling effect parameters, the sub-model selection coefficient Q is obtained; Q general decoupling sub-models, including the exclusive decoupling sub-model and the general decoupling model other than the exclusive decoupling sub-model, are selected to decouple the load temperature of the difficult-to-decouple user group, and exclusive decoupling results and general decoupling results are obtained. The general decoupling result is the average of the Q candidate decoupling results of the general decoupling sub-models. The load-temperature decoupling result is obtained by weighting the specific decoupling result and the general decoupling result. The weights in the weighting calculation are obtained based on industry temperature sensitivity parameters.
10. A load-temperature decoupling system based on user characteristic analysis, characterized in that, For implementing the load-temperature decoupling method based on user feature analysis as described in any one of claims 1-9, the system comprises: The preliminary decoupling module is used to obtain a general decoupling model, perform load-temperature decoupling based on the current temperature and current user information, and obtain preliminary decoupling results. The decoupling effect analysis module is used to obtain the decoupling residual based on the preliminary decoupling result, and to correct the decoupling residual based on the industry temperature sensitivity parameter and power capacity fluctuation rate in the current user information, so as to obtain the decoupling effect parameter. The dedicated model acquisition module is used to add users whose decoupling effect parameters are less than the decoupling effect threshold into the difficult-to-decoupling user group. Based on the user characteristics, historical temperature and historical load of the difficult-to-decoupling user group, the best general decoupling sub-model is selected for incremental learning to obtain the dedicated decoupling sub-model. The load-temperature decoupling module is used to perform load-temperature decoupling for difficult-to-decouple user groups by employing the dedicated decoupling sub-model and combining it with a general decoupling model other than the dedicated decoupling sub-model, and to obtain the load-temperature decoupling results.