A method for evaluating a user resource dynamic adjustment capability

Abstract

This application discloses a method for evaluating the dynamic adjustment capability of user resources, comprising: acquiring real-time resource data, inputting it into a prediction model for the dynamic adjustment capability of user resources, and outputting predicted data. The prediction model is trained on a dual-drive model of mechanism data under the constraint of user resource adjustment mechanism. The dual-drive model includes convolutional layers, a feedforward neural network, a multi-head attention module, a Transformer encoder, and a fusion module that fuses local features with temporal features. The predicted data is input into an evaluation model coupled with indicators, and the model outputs quantitative evaluation results on adjustment accuracy, response speed, and duration of continuous adjustment. Therefore, the dual-drive model, relying on convolutional layers and a fusion module to simultaneously extract local operational features and multi-scale temporal patterns of resources, improves the accuracy of adjustment quantity prediction. The three-dimensional indicator model accurately quantifies the user's true adjustable potential, contributing to flexible power grid dispatch and safe and stable operation.

Description

Technical Field

[0001] This application relates to the field of power information technology, and more specifically, to a method for evaluating the dynamic adjustment capability of user resources. Background Technology

[0002] As the peak-valley load difference in the power grid continues to widen, the scale of distributed power generation is constantly growing, and the volume of user-side adjustable resources such as energy storage, charging piles, and flexible loads is rapidly expanding, the demand for accurate quantitative assessment of the grid dispatching's ability to dynamically adjust user resources is continuously increasing. Existing assessment methods mostly rely on mechanistic modeling or purely data-driven modeling, which are difficult to adapt to the needs of coordinated control of multiple types of users and dispatching in multiple scenarios. They also have significant shortcomings in prediction accuracy, time-series feature capture, and high-dimensional index processing, making it difficult to fully tap the adjustment potential of user-side resources.

[0003] Traditional assessment schemes, on the one hand, cannot integrate the physical mechanisms of regulation with the characteristics of time-series data, making it difficult to characterize the temporal correlation of resource regulation and resulting in low accuracy in predicting regulation amounts; on the other hand, the use of conventional TOPSIS and PCA to process assessment indicators cannot effectively eliminate indicator redundancy, resulting in insufficient assessment robustness and computational efficiency, and failing to accurately quantify the user's actual regulation capabilities.

[0004] Therefore, how to assess the dynamic adjustment capability of user resources, and improve the accuracy of the assessment of user resource adjustment potential by increasing the prediction accuracy of user resource adjustment capability, thereby supporting the safe and stable operation and flexible dispatch of the power grid, is an issue that needs attention. Summary of the Invention

[0005] In view of the above problems, this application provides a method for evaluating the dynamic adjustment capability of user resources, so as to improve the accuracy of the evaluation of the adjustment potential of user resources by improving the prediction accuracy of the user resource adjustment capability, and support the safe and stable operation and flexible dispatch of the power grid.

[0006] To achieve the above objectives, the following specific solutions are proposed:

[0007] A method for evaluating the dynamic adjustment capability of user resources, comprising:

[0008] Obtain real-time resource data of the target user;

[0009] The real-time resource data is input into a pre-trained prediction model for the dynamic adjustment capability of user resources, and the predicted data is output. The prediction model for the dynamic adjustment capability of user resources is trained on a dual-drive model of mechanism data under the constraint of user resource adjustment mechanism. The dual-drive model of mechanism data includes a convolutional layer, a feedforward neural network, a multi-head attention module, a Transformer encoder, and a fusion module for fusing local features and temporal features.

[0010] The predicted data is input into a pre-built index coupling evaluation model, which outputs quantitative evaluation results of the real-time resource data on adjustment accuracy, response speed and continuous adjustment duration. The index coupling evaluation model couples the adjustment accuracy, response speed and continuous adjustment duration.

[0011] Optionally, the mechanism data dual-drive model extracts features using the following formula:

[0012]

[0013] in, The feature vector output by the dual-drive model of the mechanism data at time t. For the convolutional layer, The input data is at time t. For the multi-head attention module, This refers to the feedforward neural network.

[0014] Optionally, the training process of the user resource dynamic adjustment capability prediction model includes:

[0015] Obtain user resource data;

[0016] The user resource data is divided according to the type of operating scenario to obtain resource data under various operating scenarios;

[0017] For resource data in each operating scenario, the operating boundary threshold of the resource data is determined, and the target parameter associated with the adjustment amount is screened by the Pearson correlation coefficient method. Based on the resource data in the operating scenario, the operating boundary threshold, and the target parameter, the user resource adjustment business mechanism information in the operating scenario is determined.

[0018] Based on the user resource adjustment business mechanism information in various operating scenarios, and the target parameters in all operating scenarios, a Bayesian mechanism prediction model is constructed.

[0019] Using a portion of the user resource data as training samples, the mechanism data dual-drive model is iteratively trained through backpropagation algorithm under the dual-drive fusion loss function to obtain the user resource dynamic adjustment capability prediction model. The dual-drive fusion loss function includes a mechanism constraint loss term, which is constructed based on the physical constraints and adjustment boundary conditions of the Bayesian mechanism prediction model.

[0020] Optionally, the step of screening the target parameters associated with the adjustment amount using the Pearson correlation coefficient method includes:

[0021] The Pearson correlation coefficients of the parameter pairs formed by various parameters in the resource data for each operating scenario are calculated using the following formula:

[0022]

[0023] in, The parameter pair consisting of parameter a and parameter b Pearson correlation coefficient, Let a be the k-th sample value of parameter a. Let b be the k-th sample value. Let a be the sample mean of parameter a. Let b be the sample mean of parameter b. Indicates the total sample size;

[0024] Parameter pairs whose Pearson correlation coefficients are higher than a preset correlation score threshold are identified as target parameters associated with the adjustment amount.

[0025] Optionally, after acquiring the user resource data, the method further includes:

[0026] The isolated score for each sample in the user resource data is calculated using the following formula:

[0027]

[0028] in, Indicates the target sample. Indicates the number of samples. To score the isolated data of the target sample, For the target sample Path length in an isolated tree For the target sample Average path length across all isolated trees This is the average path length correction factor for the given sample size;

[0029] The user resource data is filtered and removed if the isolated data score is greater than a preset abnormal score threshold, resulting in cleaned user resource data.

[0030] Optionally, the prediction mechanism of the Bayesian mechanistic prediction model is as follows:

[0031]

[0032] in, This is the predicted value of the user resource adjustment amount at time t+1. For the Bayesian mechanism prediction model, This represents the original sequence of user resource adjustments from time 1 to time t. These are the model parameters of the Bayesian mechanism prediction model. To predict the probability distribution, Let be the posterior distribution of the model parameters.

[0033] Optionally, the dual-drive fusion loss function is:

[0034]

[0035] in, This represents the predicted value during the training phase. The mechanistic constraint loss value is calculated based on the predicted values ​​from the training phase. For data loss values, This represents the total loss value.

[0036] The mechanism data dual-driven model is trained iteratively using the backpropagation algorithm based on the following formula:

[0037]

[0038] in, These are the model parameters after the (k+1)th iteration. These are the model parameters after the k-th iteration. The adaptive learning rate for the k-th iteration. This is the gradient of the dual-drive fusion loss function at the k-th iteration.

[0039] Optionally, the method further includes:

[0040] After training the dual-drive model based on the mechanistic data, the model parameters of the dual-drive model based on the mechanistic data are updated using the following formula:

[0041]

[0042] in, The updated model parameters are those of the dual-drive model based on the aforementioned mechanism data. These are the current model parameters of the dual-drive mechanism data model. To improve the inertia weight in the particle swarm optimization algorithm, To improve the particle update speed of the particle swarm optimization algorithm, The learning rate for the Adam algorithm. This represents the gradient of the dual-drive fusion loss function under the current model parameters.

[0043] Optionally, the predicted data is input into a pre-built index-coupled evaluation model, which outputs quantitative evaluation results of the real-time resource data on adjustment accuracy, response speed, and continuous adjustment duration, including:

[0044] The predicted data is input into a pre-built index coupling evaluation model, which generates an evaluation index matrix based on the predicted data, which is coupled with the adjustment accuracy index, the response speed index, and the continuous adjustment duration index.

[0045] The dimension-reduced index matrix of the target user's real-time resource data is obtained by using the variational autoencoder of the index coupling evaluation model to reduce the dimension of the evaluation index matrix.

[0046] For each evaluation index among the adjustment accuracy index, the response speed index, and the continuous adjustment duration index, determine the optimal and worst values ​​of the resource data for each user in the evaluation index.

[0047] The index value of the dimensionality-reduced index matrix in each evaluation index is calculated using the index coupling evaluation model, and a first distance value between the index value in each evaluation index and the optimal value, and a second distance value between the index value in each evaluation index and the worst value are calculated.

[0048] For each evaluation indicator among the adjustment accuracy indicator, the response speed indicator, and the continuous adjustment duration indicator, the first distance value and the second distance value of the evaluation indicator are closely approximated and calculated using the indicator coupling evaluation model to obtain the closeness of the real-time resource data on the evaluation indicator.

[0049] Optionally, for each evaluation indicator among the adjustment accuracy indicator, the response speed indicator, and the continuous adjustment duration indicator, the first distance value and the second distance value of the evaluation indicator are calculated using the indicator coupling evaluation model to obtain the closeness of the real-time resource data on the evaluation indicator, including:

[0050] The closeness of the real-time resource data to each evaluation metric is calculated using the following formula:

[0051]

[0052] in, The degree of closeness of the real-time resource data on the i-th evaluation metric. The first distance value corresponding to the i-th evaluation index. The second distance value corresponds to the i-th evaluation index.

[0053] Using the above technical solution, this application obtains real-time resource data of the target user, inputs the real-time resource data into a pre-trained predictive model of user resource dynamic adjustment capability, and outputs predicted data. The predictive model of user resource dynamic adjustment capability is trained on a mechanism data dual-drive model under the constraint of user resource adjustment mechanism. The mechanism data dual-drive model includes convolutional layers, feedforward neural networks, multi-head attention modules, Transformer encoders, and fusion modules for fusing local features and temporal features. The predicted data is input into a pre-constructed index coupling evaluation model, and the model outputs quantitative evaluation results of real-time resource data on adjustment accuracy index, response speed index, and continuous adjustment duration index. The index coupling evaluation model couples the adjustment accuracy index, response speed index, and continuous adjustment duration index. It is evident that the dual-drive model, relying on convolutional layers, multi-head attention, and temporal fusion modules to simultaneously extract local operational features and multi-scale temporal patterns of resources, improves the accuracy of regulation prediction. By evaluating the model through a three-dimensional index of coupled regulation accuracy, response speed, and continuous regulation duration, it can accurately quantify the user's real adjustable potential, significantly improve the evaluation of robustness and computational efficiency, and provide a reliable quantitative basis for the flexible scheduling and safe and stable operation of the power grid. Attached Figure Description

[0054] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0055] Figure 1 This is a schematic diagram of a process for evaluating the ability to dynamically adjust user resources, provided in an embodiment of this application.

[0056] Figure 2 A schematic diagram of the training process for a user resource dynamic adjustment capability prediction model provided in an embodiment of this application;

[0057] Figure 3 This is a flowchart illustrating a multi-index quantitative evaluation of the mechanical energy of adjustable prediction results for user resources, provided as an embodiment of this application. Detailed Implementation

[0058] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a computer, cloud, or server.

[0060] Next, combined Figure 1 The method for evaluating the user resource dynamic adjustment capability described in this application may include the following steps:

[0061] Step S110: Obtain real-time resource data of the target user.

[0062] Real-time resource data may include real-time operating power of user-side equipment, equipment adjustment response time, real-time charging and discharging status of energy storage, charging pile load power, flexible load start-stop adjustment amount, real-time dispatch instructions issued by the power grid, real-time temperature and humidity of the site environment, real-time energy consumption load curve of users, real-time equipment operation fault alarm data, and resource adjustment execution feedback data.

[0063] Specifically, it can collect real-time resource data from the target user side across all dimensions and time series, covering the entire operation process of various adjustable resources of the user, and continuously and synchronously collecting real-time information related to the real-time operating status of equipment, interaction commands, external environment, and user energy consumption behavior. During the collection process, basic data standardization is completed simultaneously, unifying data sampling frequency and data format, ensuring the alignment of timestamps from multiple data sources, and avoiding data time series misalignment caused by different devices and collection terminals. The collection process takes into account the differentiated data collection logic of various types of user resources such as energy storage, charging piles, and flexible loads, adapting corresponding collection thresholds for different devices, filtering out null value fragments caused by transmission interruptions, and completely retaining the real-time values ​​of dynamic changes during resource adjustment. No manual secondary recording and calibration are required; the data can be directly used for subsequent cleaning and prediction processes, completely reconstructing the real-time adjustment behavior of user resources and accurately reflecting the adjustable potential boundaries of user resources at different times.

[0064] Step S120: Input real-time resource data into the pre-trained user resource dynamic adjustment capability prediction model and output prediction data.

[0065] The user resource dynamic adjustment capability prediction model is trained on a dual-drive model based on mechanistic data under the constraint of user resource adjustment mechanism. This dual-drive model includes convolutional layers, a feedforward neural network, a multi-head attention module, a Transformer encoder, and a fusion module for fusing local and temporal features. Features can be extracted using the following formula:

[0066]

[0067] in, The feature vector output by the dual-drive model of the mechanism data at time t. The convolutional layer is used to extract local features from user resource adjustment data. The input data is at time t. For the multi-head attention module, The feedforward neural network is combined with a Transformer encoder to enhance its temporal feature modeling capabilities, making it suitable for multi-regional user resource regulation scenarios.

[0068] Specifically, convolutional layers (such as those using 3x3 convolutional kernels) are used to mine local spatial features in real-time user resource data. Fixed-size convolutional kernels slide through time-series data segments to extract local correlation features between adjacent time points, such as energy storage charging and discharging power, charging pile load, and flexible load adjustment. For the differentiated operating patterns of various types of adjustable resources, convolutional layers can accurately capture local details such as equipment adjustment amplitude and sudden fluctuations within short time windows, filtering out irrelevant noise interference and compressing the original high-dimensional resource data into local feature vectors. Convolutional layers provide refined local feature inputs for time-series modeling, ensuring that the model fully perceives the short-term dynamic adjustment characteristics of equipment.

[0069] Feedforward neural networks can be deployed after multi-head attention modules to perform nonlinear mapping and feature dimensionality enhancement on the fused feature vectors, strengthening the nonlinear correlation between user resource adjustment features. Through a multi-layer fully connected structure, feedforward neural networks deeply process the temporal features of the attention output, distinguishing the weight differences of different feature dimensions such as adjustment precision and response time, amplifying effective adjustment features and weakening ineffective and redundant features. Simultaneously, feedforward neural networks can adapt to the computational requirements of mechanistic constraint loss functions, complete feature standardization output, connect to the model loss backpropagation process, balance the representation weights of local and temporal features, and improve the model's ability to fit user adjustment behaviors in multiple scenarios.

[0070] The multi-head attention module can decompose input features into multiple independent attention heads for parallel computation, modeling the dependencies of user resource adjustment data at different time scales and quantifying the influence weight of historical adjustment sequences on current adjustment behavior. The multi-head attention module can simultaneously capture short-term instantaneous adjustment correlations and long-term energy consumption change patterns, identifying time-series lag characteristics such as energy storage charging and discharging lags and load adjustment response delays. The multi-head attention module can autonomously select key historical time-series samples, weakening interference from irrelevant period data to accurately characterize the time-series autocorrelation patterns of various types of adjustable resources, thus improving the accuracy of time-series dependency modeling.

[0071] The Transformer encoder is the core unit for temporal modeling. It connects a multi-head attention module with a feedforward neural network to progressively model the long-term temporal regulation patterns of user resources. The Transformer encoder introduces temporal position encoding to mark the time sequence of data, fully preserving the time-series information of real-time resource data, and is suitable for business scenarios of continuous adjustment across time periods in energy storage and charging piles. The stacking of multiple encoders can extract deep temporal features layer by layer, and combine the constraints of user resource adjustment mechanisms to correct temporal prediction biases, unifying and integrating adjustment sequence information from multiple devices and time periods to output a complete temporal feature representation.

[0072] The fusion module connects the local features output from the convolutional layer with the global temporal features output from the Transformer encoder, achieving deep coupling between the two types of features through a weighted fusion mechanism. The fusion module can dynamically adjust the fusion weights of local and temporal features based on user resource adjustment mechanism constraints, taking into account both the instantaneous adjustment details of devices and long-term operational change patterns. After fusion, a unified comprehensive feature vector is generated, simultaneously carrying local device characteristics and global temporal patterns, improving the model's accuracy in depicting the dynamic adjustment behavior of multiple types of user resources and optimizing the subsequent adjustment amount prediction output.

[0073] Understandably, by integrating the advantages of local feature extraction from convolutional layers with temporal modeling from Transformers, optimizing the structure of convolutional layers and encoders, the model's ability to capture the spatiotemporal features of user resource adjustment data is enhanced, thereby achieving accurate capture of adjustment patterns across multiple time scales.

[0074] Step S130: Input the predicted data into the pre-built index coupling evaluation model and output the quantitative evaluation results of real-time resource data on the adjustment accuracy index, response speed index and continuous adjustment duration index.

[0075] Among them, the index coupling evaluation model couples the adjustment accuracy index, the response speed index, and the continuous adjustment duration index.

[0076] Specifically, the adjustment accuracy index measures the deviation between the user resource adjustment amount predicted by the fusion model and the actual adjustment amount; its value is the proportion of the difference between the predicted and actual data to the actual data. The response speed index measures how quickly user resources complete adjustment actions after receiving adjustment commands, characterizing the ability of user resources to stably maintain the target adjustment state. The duration of continuous adjustment index measures the ability of user resources to stably maintain the target adjustment state; its value is the duration from when the resource reaches the target adjustment state until when the resource leaves the target adjustment state, for example, representing the ability to maintain user resource adjustment accuracy under extreme environments (such as sudden load fluctuations or severe weather).

[0077] The method for evaluating the dynamic adjustment capability of user resources provided in this embodiment acquires real-time resource data of the target user, inputs the real-time resource data into a pre-trained prediction model for the dynamic adjustment capability of user resources, and outputs prediction data. The prediction model for the dynamic adjustment capability of user resources is trained on a mechanism data dual-drive model under the constraint of user resource adjustment mechanism. The mechanism data dual-drive model includes convolutional layers, feedforward neural networks, multi-head attention modules, Transformer encoders, and fusion modules for fusing local features and temporal features. The prediction data is input into a pre-constructed index coupling evaluation model, and the model outputs quantitative evaluation results of real-time resource data on adjustment accuracy index, response speed index, and continuous adjustment duration index. The index coupling evaluation model couples the adjustment accuracy index, response speed index, and continuous adjustment duration index. It is evident that the dual-drive model, relying on convolutional layers, multi-head attention, and temporal fusion modules to simultaneously extract local operational features and multi-scale temporal patterns of resources, improves the accuracy of regulation prediction. By evaluating the model through a three-dimensional index of coupled regulation accuracy, response speed, and continuous regulation duration, it can accurately quantify the user's real adjustable potential, significantly improve the evaluation of robustness and computational efficiency, and provide a reliable quantitative basis for the flexible scheduling and safe and stable operation of the power grid.

[0078] In some embodiments of this application, the training process of the user resource dynamic adjustment capability prediction model mentioned in the above embodiments is described, with reference to... Figure 2 The training process may include:

[0079] Step S210: Obtain user resource data.

[0080] Specifically, it can collect multi-dimensional, multi-source user resource data covering various adjustable resources such as energy storage devices, charging piles, and flexible loads. This includes real-time and historical time-series data such as equipment operating power, adjustment response time, ambient temperature and humidity, power grid dispatch instructions, and energy load curves. It unifies data timestamps, completes basic format standardization, removes obvious transmission null values, and forms a complete, time-aligned original dataset.

[0081] Furthermore, after acquiring user resource data, the user resource data can be cleaned to obtain cleaned user resource data. The specific process may include:

[0082] The isolated score for each sample in the user resource data is calculated using the following formula:

[0083]

[0084] in, Indicates the target sample. Indicates the number of samples. The data in the target sample is isolated and scored; the closer the score is to 1, the more likely it is to be an outlier. For target samples Path length in an isolated tree For target samples Average path length across all isolated trees This is the average path length correction factor for the sample size.

[0085] It's understandable that an isolation tree is a binary tree structure built based on a random segmentation strategy. Its core physical meaning is that anomalous data points, due to their sparse distribution and features deviating from the normal range, are more easily and quickly isolated to shallow leaf nodes during random feature selection and threshold segmentation. Normal data points, on the other hand, due to their dense distribution and similar features, require multiple segmentations to be separated. For user resource adjustment data, the isolation tree construction process is as follows: A feature dimension is randomly selected from the collected multi-source data (including features such as device operating power, adjustment response time, and environmental temperature and humidity). A segmentation threshold is randomly selected within the range of that dimension's values. Samples are recursively segmented into left and right subtrees until a node contains only a single sample or reaches a preset tree height. By constructing multiple independent isolation trees to form an isolated forest, the average path length of a sample across all trees can be used to quantify its anomaly level. The improvement lies in introducing an adaptive sampling strategy, dynamically adjusting the sampling ratio based on the sample distribution characteristics of user resource types (flexible loads, energy storage devices, charging piles) to avoid anomaly detection bias caused by insufficient samples of a certain type of equipment.

[0086] Based on this, samples in the user resource data whose data isolation scores are greater than a preset abnormal score threshold can be filtered and removed to obtain cleaned user resource data.

[0087] Understandably, by calculating isolated scores for samples in the data, user resource data is filtered and cleaned to avoid anomaly detection bias caused by sample imbalance, reduce redundant calculations, detect anomalies in data collected from multiple sources, and remove invalid data caused by equipment failures and data transmission errors, thus ensuring data quality.

[0088] Furthermore, after cleaning the user resource data, the data can be divided into training, validation, and test sets in a 7:2:1 ratio to provide standardized data support for subsequent model training, validation, and testing.

[0089] Step S220: Divide user resource data according to the type of running scenario to obtain resource data under various running scenarios.

[0090] Specifically, the original user resource data can be segmented into multiple scenarios based on three dimensions: adjustable resource type, power grid dispatch period, and external environmental conditions. This distinguishes the independent operating scenarios corresponding to three types of resources: flexible load, energy storage, and charging piles. It also differentiates between peak, flat, and valley dispatch periods, and overlays high and low temperature environment scenario labels to assign a unique scenario identifier to each data point. The dataset is then categorized and aggregated based on these scenario labels, eliminating invalid data fragments that are mixed across scenarios or lack scene labels. This results in multiple independent categorized datasets with single scenario characteristics, ensuring that the regulation mechanisms under each scenario can be independently and accurately analyzed.

[0091] Step S230: For resource data under each operating scenario, determine the operating boundary threshold of the resource data, screen the target parameters associated with the adjustment amount using the Pearson correlation coefficient method, and determine the user resource adjustment business mechanism information under the operating scenario based on the resource data, operating boundary threshold and target parameters under the operating scenario.

[0092] Among them, the process of determining the operating boundary threshold of resource data for each operating scenario can be combined with the factory rated parameters of various equipment such as energy storage, charging piles, and flexible loads and the control and constraint standards issued by the power grid dispatch. The upper and lower limits of power, continuous adjustment duration, response delay and other constraint ranges are defined for each scenario. The numerical range of normal operation of resources under each scenario is sorted out, and it is clear that exceeding the range is an abnormal operating condition.

[0093] The process of screening target parameters associated with the moderating variable using the Pearson correlation coefficient method may include:

[0094] The Pearson correlation coefficients of the parameter pairs formed by various parameters in the resource data for each operating scenario are calculated using the following formula:

[0095]

[0096] in, The parameter pair consisting of parameter a and parameter b Pearson correlation coefficient, Let a be the k-th sample value of parameter a. Let b be the k-th sample value. Let a be the sample mean of parameter a. Let b be the sample mean of parameter b. This represents the total sample size. Parameters a and b are user resource adjustment-related parameters determined after scenario decomposition and constraint extraction. Specifically, they may include energy storage device charging and discharging power, flexible load adjustment response, ambient temperature, and grid dispatch command values. The linear correlation between the two is quantified using the Pearson correlation coefficient to select core parameters with strong correlation to the adjustment amount.

[0097] Based on this, parameter pairs with Pearson correlation coefficients higher than a preset correlation score threshold can be identified as target parameters associated with the adjustment amount.

[0098] Based on resource data, operational boundary thresholds, and target parameters in the operational scenario, the process of determining the user resource regulation business mechanism information in the operational scenario can combine the physical operation rules of equipment in each scenario with the grid regulation constraints, use the strongly correlated target parameters obtained by screening, and combine the defined operational boundary thresholds to sort out the coupling relationship between parameters, extract the regulation and change laws of energy storage, flexible loads, and charging piles, and form complete and quantifiable user resource regulation business mechanism information in the corresponding scenario.

[0099] Step S240: Based on the user resource adjustment business mechanism information of various operating scenarios and the target parameters under all operating scenarios, construct a Bayesian mechanism prediction model.

[0100] The prediction mechanism of the Bayesian mechanistic prediction model is as follows:

[0101]

[0102] in, This is the predicted value of the user resource adjustment amount at time t+1. For Bayesian mechanism prediction models, This represents the original sequence of user resource adjustments from time 1 to time t. These are the model parameters for the Bayesian mechanistic prediction model. To predict the probability distribution, Let be the posterior distribution of the model parameters.

[0103] Specifically, the Bayesian mechanistic prediction model can introduce user resource regulation lag factors and autocorrelation attention mechanisms. By quantifying the influence weights of different regulation parameters on user resource regulation behavior, the model structure and parameter mapping relationship can be optimized, thereby improving the accuracy of the mechanistic model in depicting the user resource regulation patterns across multiple time scales.

[0104] Step S250: Using a portion of the user resource data as training samples, the mechanism data dual-drive model is iteratively trained through backpropagation algorithm under the dual-drive fusion loss function to obtain the predictive model of user resource dynamic adjustment capability.

[0105] The dual-drive fusion loss function includes a mechanistic constraint loss term. This mechanistic constraint loss term is constructed based on the physical constraints and adjusted boundary conditions of the Bayesian mechanistic prediction model. Specifically, the dual-drive fusion loss function is as follows:

[0106]

[0107] in, This represents the predicted value during the training phase. This is the mechanistic constraint loss value calculated based on the predicted values ​​during the training phase. For data loss values, This represents the total loss value.

[0108] The mechanism-data dual-driven model can be iteratively trained using the backpropagation algorithm using the following formula:

[0109]

[0110] in, These are the model parameters after the (k+1)th iteration. These are the model parameters after the k-th iteration. is the adaptive learning rate for the k-th iteration (the initial value can be set to 0.001, and it decays by 10% every 100 iterations). This is the gradient of the dual-drive fusion loss function at the k-th iteration.

[0111] Furthermore, a hybrid optimization algorithm combining an improved particle swarm optimization algorithm and the Adam algorithm can be used to optimize the prediction model of user resource dynamic adjustment capability. Specifically, based on the deviation found by comparing the predicted adjustment amount with the actual adjustment amount, the number of improved particles and the penalty factor can be dynamically adjusted to optimize the core parameters, mechanism constraint weights, and data preprocessing logic of the prediction model of user resource dynamic adjustment capability. The specific process may include:

[0112] After training the mechanistic data-driven dual-drive model, the model parameters are updated using the following formula:

[0113]

[0114] in, The updated model parameters are for the dual-driven model based on mechanistic data. The current model parameters are for the dual-driven model of mechanism and data. To improve the inertia weights of the particle swarm optimization algorithm, this is used to dynamically adjust the particle search step size. To improve the particle update speed of the particle swarm optimization algorithm, This is the learning rate for the Adam algorithm (e.g., set to 0.001), used to control the step size for parameter updates. This represents the gradient of the dual-drive fusion loss function under the current model parameters. The number of particles is set to 30, and the penalty factor is set to 0.01.

[0115] The method for evaluating the dynamic adjustment capability of user resources provided in this embodiment overcomes the shortcomings of traditional mechanism modeling, which cannot accurately depict the timing pattern of adjustment and has unclear parameter correlation, by conducting an analysis of the user resource adjustment mechanism, constructing an improved Bayesian mechanism prediction model, and introducing a adjustment lag factor and an autocorrelation attention mechanism, thus achieving accurate prediction of the user resource adjustment amount.

[0116] Furthermore, by constructing a dual-driven fusion model of mechanism and data, combined with Bayesian mechanism constraints, a deep coupling between regulation mechanism and data features is achieved. This overcomes the shortcomings of traditional single-mechanism modeling based on physical formula derivation and single-data-driven modeling relying solely on historical data, which leads to a disconnect between mechanism and data. It accurately captures the temporal and local features of user resource regulation, further improving the evaluation accuracy.

[0117] In some embodiments of this application, the process of step S130 above, which involves inputting predicted data into a pre-built index-coupled evaluation model and outputting quantitative evaluation results of real-time resource data on adjustment accuracy, response speed, and continuous adjustment duration, is described. (Refer to...) Figure 3 The process may include:

[0118] Step S310: Input the predicted data into the pre-built index coupling evaluation model. Based on the predicted data, the index coupling evaluation model generates an evaluation index matrix coupled with the adjustment accuracy index, response speed index, and continuous adjustment duration index.

[0119] Specifically, the predicted user resource adjustment amount output by the user resource dynamic adjustment capability prediction model can be standardized and uniformly processed to calculate the three core indicator values ​​of adjustment deviation, adjustment response time, and continuous adjustable duration for the corresponding samples. The three indicator values ​​are then integrated and arranged according to the sample dimensions to construct a standardized evaluation indicator matrix that is dimension-aligned and coupled with indicators, fully reflecting the multi-dimensional adjustment capability characteristics of the target user.

[0120] Step S320: Reduce the dimensionality of the evaluation index matrix by using the variational autoencoder of the index coupling evaluation model to obtain the dimensionality-reduced index matrix of the target user's real-time resource data.

[0121] Specifically, the high-dimensional coupled index matrix can be compressed by calling the built-in encoder module of the variational autoencoder. At the same time, redundant correlations and multicollinearity interference between indices are eliminated, while the core effective features of adjustment accuracy, response speed and continuous adjustment duration are retained, and a low-dimensional, non-redundant standardized dimensionality-reduced index matrix is ​​output.

[0122] Step S330: For each evaluation index among the adjustment accuracy index, response speed index, and continuous adjustment duration index, determine the optimal and worst values ​​of the resource data for each user in the evaluation index.

[0123] The optimal value represents the best performance of each of the three indicators—adjustment accuracy, response speed, and continuous adjustment duration—among all user samples. This constitutes the positive ideal solution for evaluation and represents the theoretically highest level of user resource adjustment capability. The worst value represents the worst performance of each of the three indicators among all user samples. This constitutes the negative ideal solution for evaluation and represents the lowest baseline level of user resource dynamic adjustment capability, serving as a benchmark for comparison.

[0124] Specifically, the entire set of user indicator samples after dimensionality reduction using a variational autoencoder can be traversed, and three independent indicator sequences—adjustment accuracy, response speed, and continuous adjustment duration—can be separated. Numerical selection is then performed separately for each indicator category. For the adjustment accuracy indicator, the value with the smallest adjustment deviation is selected as the optimal value, and the value with the largest deviation is selected as the worst value. For the response speed indicator, the shortest adjustment response time is selected as the optimal value, and the longest response time is selected as the worst value. For the continuous adjustment duration indicator, the longest adjustable duration is selected as the optimal value, and the shortest adjustable duration is selected as the worst value. The optimal and worst values ​​corresponding to the three types of indicators are stored uniformly, constructing a complete positive and negative ideal solution benchmark. This provides a unified reference standard for subsequent distance calculations and proximity metric evaluations, ensuring that the benchmark for comparing the adjustment capabilities of all users is consistent and eliminating evaluation bias caused by differences in indicator magnitude.

[0125] Step S340: Calculate the index value of the dimensionality-reduced index matrix in each evaluation index through the index coupling evaluation model, and calculate the first distance value between the index value in each evaluation index and the optimal value, and the second distance value between the index value in each evaluation index and the worst value.

[0126] The first distance value represents the Euclidean distance between the sample of dimensionality-reduced indicators of the user under evaluation and the positive ideal solution formed by the optimal values ​​of each indicator. It quantifies the gap between the user's adjustment capability and the ideal optimal adjustment level. The smaller the distance value, the closer the user's overall performance in adjustment accuracy, response speed, and continuous adjustment duration is to the optimal standard, and the more prominent the adjustment potential. It is a core positive reference indicator for judging the quality of a user's adjustability. The second distance value represents the Euclidean distance between the sample of dimensionality-reduced indicators of the user under evaluation and the negative ideal solution formed by the worst values ​​of each indicator. It measures the gap between the user's adjustment capability and the worst baseline level. The larger the distance value, the lower the probability that the user's various adjustment indicators are far worse than the worst standard, indicating better adjustability. It can be used in conjunction with the first distance value to complete the quantitative calculation of overall closeness.

[0127] Step S350: For each evaluation indicator among the adjustment accuracy indicator, response speed indicator, and continuous adjustment duration indicator, the first distance value and the second distance value of the evaluation indicator are closely quantified and calculated through the indicator coupling evaluation model to obtain the closeness of the real-time resource data on the evaluation indicator.

[0128] Specifically, the process of calculating the closeness of real-time resource data to each evaluation metric may include:

[0129] The following formula is used to calculate the closeness of real-time resource data to each evaluation metric:

[0130]

[0131] in, To determine the closeness of real-time resource data to the i-th evaluation metric, Let be the first distance value corresponding to the i-th evaluation index. This is the second distance value corresponding to the i-th evaluation index.

[0132] Understandably, proximity is a core quantitative indicator that quantifies the comprehensive ability of a target user to dynamically adjust resources, ranging from 0 to 1. Proximity comprehensively measures the relative distance between the sample being evaluated and the optimal adjustment level and the worst adjustment benchmark. The closer the value is to 1, the better the user's overall performance in terms of adjustment accuracy, response speed, and duration of continuous adjustment; the closer the value is to 0, the worse the adjustment ability.

[0133] The user resource dynamic adjustment capability assessment method provided in this embodiment establishes a scientific assessment indicator system through an indicator coupling assessment model. It reduces the dimensionality and redundancy of high-dimensional assessment indicators and optimizes the assessment logic, effectively overcoming the shortcomings of traditional technical indicators such as redundancy, low assessment efficiency, and poor robustness, and achieving a dual improvement in assessment efficiency and robustness.

[0134] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for evaluating the dynamic adjustment capability of user resources, characterized in that, include: Obtain real-time resource data of the target user; The real-time resource data is input into a pre-trained prediction model for the dynamic adjustment capability of user resources, and the predicted data is output. The prediction model for the dynamic adjustment capability of user resources is trained on a dual-drive model of mechanism data under the constraint of user resource adjustment mechanism. The dual-drive model of mechanism data includes a convolutional layer, a feedforward neural network, a multi-head attention module, a Transformer encoder, and a fusion module for fusing local features and temporal features. The predicted data is input into a pre-built index coupling evaluation model, which outputs quantitative evaluation results of the real-time resource data on adjustment accuracy, response speed and continuous adjustment duration. The index coupling evaluation model couples the adjustment accuracy, response speed and continuous adjustment duration.

2. The method according to claim 1, characterized in that, The mechanism data dual-drive model extracts features using the following formula: in, The feature vector output by the dual-drive model of the mechanism data at time t. For the convolutional layer, The input data is at time t. For the multi-head attention module, This refers to the feedforward neural network.

3. The method according to claim 1, characterized in that, The training process of the user resource dynamic adjustment capability prediction model includes: Obtain user resource data; The user resource data is divided according to the type of operating scenario to obtain resource data under various operating scenarios; For resource data in each operating scenario, the operating boundary threshold of the resource data is determined, and the target parameter associated with the adjustment amount is screened by the Pearson correlation coefficient method. Based on the resource data in the operating scenario, the operating boundary threshold, and the target parameter, the user resource adjustment business mechanism information in the operating scenario is determined. Based on the user resource adjustment business mechanism information in various operating scenarios, and the target parameters in all operating scenarios, a Bayesian mechanism prediction model is constructed. Using a portion of the user resource data as training samples, the mechanism data dual-drive model is iteratively trained through backpropagation algorithm under the dual-drive fusion loss function to obtain the user resource dynamic adjustment capability prediction model. The dual-drive fusion loss function includes a mechanism constraint loss term, which is constructed based on the physical constraints and adjustment boundary conditions of the Bayesian mechanism prediction model.

4. The method according to claim 3, characterized in that, The screening of target parameters associated with the adjustment amount using the Pearson correlation coefficient method includes: The Pearson correlation coefficients of the parameter pairs formed by various parameters in the resource data for each operating scenario are calculated using the following formula: in, The parameter pair consisting of parameter a and parameter b Pearson correlation coefficient, Let a be the k-th sample value of parameter a. Let b be the k-th sample value. Let a be the sample mean of parameter a. Let b be the sample mean of parameter b. Indicates the total sample size; Parameter pairs whose Pearson correlation coefficients are higher than a preset correlation score threshold are identified as target parameters associated with the adjustment amount.

5. The method according to claim 3, characterized in that, After acquiring the user resource data, the method further includes: The isolated score for each sample in the user resource data is calculated using the following formula: in, Indicates the target sample. Indicates the number of samples. To score the isolated data of the target sample, For the target sample Path length in an isolated tree For the target sample Average path length across all isolated trees This is the average path length correction factor for the given sample size; The user resource data is filtered and removed if the isolated data score is greater than a preset abnormal score threshold, resulting in cleaned user resource data.

6. The method according to claim 3, characterized in that, The prediction mechanism of the Bayesian mechanism prediction model is as follows: in, This is the predicted value of the user resource adjustment amount at time t+1. For the Bayesian mechanism prediction model, This represents the original sequence of user resource adjustments from time 1 to time t. These are the model parameters of the Bayesian mechanism prediction model. To predict the probability distribution, Let be the posterior distribution of the model parameters.

7. The method according to any one of claims 3-6, characterized in that, The dual-drive fusion loss function is: in, This represents the predicted value during the training phase. The mechanistic constraint loss value is calculated based on the predicted values ​​from the training phase. For data loss values, This represents the total loss value. The mechanism data dual-driven model is trained iteratively using the backpropagation algorithm based on the following formula: in, These are the model parameters after the (k+1)th iteration. These are the model parameters after the k-th iteration. The adaptive learning rate for the k-th iteration. This is the gradient of the dual-drive fusion loss function at the k-th iteration.

8. The method according to claim 7, characterized in that, The method also includes: After training the dual-drive model based on the mechanistic data, the model parameters of the dual-drive model based on the mechanistic data are updated using the following formula: in, The updated model parameters are those of the dual-drive model based on the aforementioned mechanism data. These are the current model parameters of the dual-drive mechanism data model. To improve the inertia weight in the particle swarm optimization algorithm, To improve the particle update speed of the particle swarm optimization algorithm, The learning rate for the Adam algorithm. This represents the gradient of the dual-drive fusion loss function under the current model parameters.

9. The method according to any one of claims 1-8, characterized in that, The predicted data is input into a pre-built index-coupled evaluation model, which outputs quantitative evaluation results of the real-time resource data on adjustment accuracy, response speed, and continuous adjustment duration, including: The predicted data is input into a pre-built index coupling evaluation model, which generates an evaluation index matrix based on the predicted data, which is coupled with the adjustment accuracy index, the response speed index, and the continuous adjustment duration index. The dimension-reduced index matrix of the target user's real-time resource data is obtained by using the variational autoencoder of the index coupling evaluation model to reduce the dimension of the evaluation index matrix. For each evaluation index among the adjustment accuracy index, the response speed index, and the continuous adjustment duration index, determine the optimal and worst values ​​of the resource data for each user in the evaluation index. The index value of the dimensionality-reduced index matrix in each evaluation index is calculated using the index coupling evaluation model, and a first distance value between the index value in each evaluation index and the optimal value, and a second distance value between the index value in each evaluation index and the worst value are calculated. For each evaluation indicator among the adjustment accuracy indicator, the response speed indicator, and the continuous adjustment duration indicator, the first distance value and the second distance value of the evaluation indicator are closely approximated and calculated using the indicator coupling evaluation model to obtain the closeness of the real-time resource data on the evaluation indicator.

10. The method according to claim 9, characterized in that, For each evaluation indicator among the adjustment accuracy indicator, the response speed indicator, and the continuous adjustment duration indicator, the first distance value and the second distance value of the evaluation indicator are calculated using the indicator coupling evaluation model to obtain the closeness of the real-time resource data on the evaluation indicator, including: The closeness of the real-time resource data to each evaluation metric is calculated using the following formula: in, The degree of closeness of the real-time resource data on the i-th evaluation metric. The first distance value corresponding to the i-th evaluation index. The second distance value corresponds to the i-th evaluation index.

Images