Method and system for evaluating operation energy efficiency of all-electric energy station based on quantity-quality synergy

Abstract

A kind of based on quantity and quality synergistic all-electric energy station operation energy efficiency evaluation method and system, comprising: collecting original data, and carrying out data cleaning, and the energy efficiency grade label annotation is carried out to cleaned data;Original data is handled with time-space alignment, and generates time-space alignment feature tensor;Based on dynamic convolution kernel, the one-dimensional convolution operation of multi-scale is carried out to the time-space alignment feature tensor, and the convolution result of different scales is fused, and generates multi-scale feature tensor;All-electric energy station operation energy efficiency evaluation model is constructed, and the training is carried out to multi-scale feature tensor data, and the training all-electric energy station operation energy efficiency evaluation model is obtained;Real-time acquisition data is extracted after feature, and input all-electric energy station operation energy efficiency evaluation model, and the all-electric energy station operation energy efficiency evaluation result is obtained.The present application can accurately and reliably evaluate all-electric energy station operation energy efficiency, and provide strong support for energy station planning and operation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method and system for evaluating the operational energy efficiency of all-electric energy stations based on quantity-quality synergy. Background Technology

[0002] With the transformation of the energy structure and the pursuit of efficient, clean and reliable energy supply, all-electric energy stations, as a system integrating multiple power equipment and energy conversion technologies, are playing an increasingly important role in the modern energy system. However, their complex operating characteristics, involving multi-dimensional time-series data such as power, current, voltage, temperature, equipment switching status, and load rate, pose a huge challenge to the accurate assessment of the operating energy efficiency of all-electric energy stations.

[0003] Therefore, there is an urgent need for a method to evaluate the energy efficiency of all-electric energy stations with strong anti-interference capabilities and accurate assessment, so as to effectively integrate multi-dimensional time-series data, eliminate interference caused by data noise and dimensional differences, and achieve accurate quantitative assessment of the energy efficiency of all-electric energy stations, providing reliable decision support for their operation optimization, equipment maintenance and low-carbon transformation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for evaluating the operational energy efficiency of all-electric energy stations based on quantity and quality synergy.

[0005] The present invention adopts the following technical solution.

[0006] In a first aspect, the present invention provides a method for evaluating the operational energy efficiency of a full-electric energy station based on quantity-quality synergy, comprising the following steps:

[0007] S1. Collect raw data and clean the data, then label the cleaned data with energy efficiency rating labels.

[0008] S2. Perform spatiotemporal alignment processing on the original data to generate a spatiotemporal aligned feature tensor;

[0009] S3. Perform multi-scale one-dimensional convolution operation on the spatiotemporal aligned feature tensor based on dynamic convolution kernel, and fuse the convolution results of different scales to generate a multi-scale feature tensor.

[0010] S4. Construct an energy efficiency evaluation model for the operation of the all-electric energy station. The model is trained by training multi-scale feature tensor data.

[0011] S5. After processing the real-time collected data through S1-S3, input it into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation result of the all-electric energy station.

[0012] More preferably,

[0013] The raw data includes the power, current, voltage, temperature, on / off status, and load rate of each device in the all-electric power station.

[0014] More preferably,

[0015] S2 specifically includes:

[0016] For each type of equipment in the all-electric power station, the minimum dynamic time warping path distance between its original time series data and the benchmark template generated by clustering historical data is calculated. Based on this distance value and combined with the equipment influence attenuation factor, a weighting coefficient is dynamically generated. The dynamic time warping algorithm is applied to the original time series data of each type of equipment to adjust its length to the standard time length defined by the benchmark template. Based on the weighting coefficient, the aligned results of all types of equipment are weighted and fused to generate a spatiotemporal alignment feature tensor of uniform length.

[0017] More preferably,

[0018] The dynamic convolution kernel is generated in the following specific way:

[0019] Obtain the operating condition statistics of the current sample data and calculate their mean to obtain the operating condition statistics mean vector. Input the operating condition statistics mean vector into the multilayer perceptron. The multilayer perceptron calculates and outputs the convolution kernel weight scaling factor. Multiply the learnable basic convolution kernel by the convolution kernel weight scaling factor to obtain the dynamic convolution kernel.

[0020] More preferably,

[0021] The energy efficiency evaluation model for the all-electric energy station includes an input layer, a feature processing layer, and a decision layer. The input layer receives a multi-scale feature tensor and performs preliminary dimensionality reduction through a feature selector network to output a low-dimensional feature vector. The feature processing layer includes a mutual information feature compression module and a Bayesian attention module, used to calculate the final uncertainty-aware feature representation from the low-dimensional feature vector. The decision layer, based on the final uncertainty-aware feature representation and combined with evidence theory, generates the probability of the energy efficiency evaluation category and the classification confidence level corresponding to the current input sample. It then uses negative log-likelihood to calculate the classification loss and calculates the classification loss function by using the difference between the Dirichlet distribution and the uniform distribution with KL divergence penalty parameter as the evidence value.

[0022] More preferably,

[0023] The mutual information feature compression module specifically includes:

[0024] The mutual information between the multi-scale feature tensor and the target energy efficiency level label is calculated. The difference between the product of the mutual information between the multi-scale feature tensor and the low-dimensional feature vector and the compression strength coefficient is subtracted as the mutual information loss function, which is used to compress redundant features while retaining high-information features.

[0025] More preferably,

[0026] The Bayesian attention module specifically includes:

[0027] A working condition query vector is generated based on the mean vector of the working condition statistics, and interacts with the low-dimensional feature vector to calculate the mean of the attention weight.

[0028] The variance of attention weights is calculated using the variance projection matrix and the ReLU activation function; the variance projection matrix is ​​a learnable parameter.

[0029] The attention weights are modeled as a Gaussian distribution, which is defined by the mean and variance of the attention weights.

[0030] Attention weights are sampled from a Gaussian distribution using a reparameterization technique. The sampled attention weights are then used to perform a weighted summation on the low-dimensional feature vector to generate the final uncertainty-aware feature representation.

[0031] More preferably,

[0032] In the energy efficiency evaluation model for the operation of the all-electric energy station, the L2 norm of the network parameter matrix of each network layer is calculated and divided by the square of the uncertainty metric of that layer as a regularization term; the regularization term is added to the mutual information loss function and the classification loss function to obtain the total loss function.

[0033] More preferably,

[0034] In S5, when the highest category probability of the energy efficiency evaluation result of the all-electric energy station is greater than the preset highest category probability threshold and the uncertainty metric value is less than the preset minimum uncertainty metric threshold, the classification result is directly adopted; when the uncertainty metric value is greater than or equal to the preset minimum uncertainty metric threshold or the difference between the probabilities of each category is less than the preset minimum probability difference threshold, the manual review mechanism is triggered.

[0035] Secondly, the present invention provides an energy efficiency evaluation system for all-electric energy stations based on the aforementioned method, including a data acquisition module, a spatiotemporal aligned feature tensor generation module, a multi-scale feature tensor generation module, an energy efficiency evaluation model construction module for all-electric energy stations, and an energy efficiency evaluation result calculation module for all-electric energy stations.

[0036] The data acquisition module is used to collect raw data, perform data cleaning, and label the cleaned data with energy efficiency level labels.

[0037] The spatiotemporal alignment feature tensor generation module is used to perform spatiotemporal alignment processing on the original data and generate a spatiotemporal alignment feature tensor.

[0038] The multi-scale feature tensor generation module is used to perform multi-scale one-dimensional convolution operations on the spatiotemporally aligned feature tensor based on dynamic convolution kernels, and fuse the convolution results of different scales to generate a multi-scale feature tensor.

[0039] The module for constructing an energy efficiency evaluation model for the operation of a full-electric energy station is used to construct an energy efficiency evaluation model for the operation of a full-electric energy station. By training multi-scale feature tensor data, the trained energy efficiency evaluation model for the operation of the full-electric energy station is obtained.

[0040] The module for calculating the energy efficiency evaluation results of the all-electric energy station is used to extract features from real-time collected data to generate multi-scale feature tensors, which are then input into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation results of the all-electric energy station.

[0041] The beneficial effects of this invention are compared with those of the prior art:

[0042] To address the problem of causal distortion caused by the lack of effective alignment methods for multi-source asynchronous time-series data in existing technologies, this invention adopts a dynamic time warping strategy to couple physical constraints of device operation to construct a spatiotemporal alignment matrix. This solves the problems of sampling frequency differences and asynchronous timestamps in multi-source sensor data, preserves the temporal correlation characteristics under physical constraints, and avoids causal distortion caused by interpolation and loss of high-frequency details in fixed-window aggregation compared to conventional methods.

[0043] To address the shortcomings of existing feature extraction methods, which struggle to adapt to scale changes caused by load fluctuations, resulting in poor high- and low-frequency feature fusion, overfitting in noisy environments, and difficulty in supporting "uncertain result" judgments, thus affecting the accuracy, stability, and interpretability of evaluations, this invention proposes a working condition-aware convolutional kernel group. This group dynamically adjusts the convolutional kernel size based on the current sample's working condition statistics, solving the scale mismatch problem between high-frequency features such as current / voltage and low-frequency features such as temperature. It adapts to working condition fluctuations such as sudden load changes and overcomes the limitation of conventional wavelet transform with its fixed preset scale, which cannot adapt to working condition fluctuations such as sudden load changes in energy stations.

[0044] In addition, by modeling the attention weights as a Gaussian distribution, the uncertainty of feature importance is quantified, and the decision confidence is quantified by variance. This avoids the problem of overfitting of attention weights in high-noise periods when conventional point estimation is used, and improves the robustness of the model under complex working conditions.

[0045] Furthermore, a subjective logical distribution is constructed based on evidence theory, and classification uncertainty is modeled through Dirichlet distribution to achieve confidence-perceived energy efficiency evaluation, explicitly quantify the uncertainty of classification decisions, support the determination of "uncertain" states, and solve the problems that conventional softmax outputs a single probability distribution that cannot quantify classification confidence and that cross-entropy loss forces low-confidence samples to misclassify.

[0046] Through the above improvements, the present invention effectively solves the problem of asynchronous sensor data. At the same time, it enables the model to maintain stable performance under sensor noise interference, avoids the defect of overfitting in conventional methods during high noise periods, and improves the model's anti-interference ability. Attached Figure Description

[0047] Figure 1 This is a flowchart of the energy efficiency evaluation method for all-electric energy stations based on quantity-quality synergy in this invention;

[0048] Figure 2 This is a comparative experimental diagram of the classification accuracy of different temporal alignment methods in Embodiment 2 of the present invention;

[0049] Figure 3 This is a comparative experimental diagram of the robustness of the method under different noise levels in Embodiment 3 of the present invention;

[0050] Figure 4 This is a comparative experimental diagram of the uncertainty measurement distribution of different methods in Embodiment 4 of the present invention;

[0051] Figure 5 This is a comparative experimental diagram showing the training convergence speed of different methods in Embodiment 5 of the present invention;

[0052] Figure 6 This is a comparative experimental diagram showing the change of key feature weights over time in Embodiment Six of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this invention are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0054] Example 1:

[0055] Reference Figure 1 This invention provides a method for evaluating the operational energy efficiency of a full-electric power station based on quantity and quality synergy, specifically including the following steps:

[0056] S1. Collect raw data and clean the data, then label the cleaned data with energy efficiency rating labels.

[0057] In response to the multi-dimensional operating characteristics of the all-electric energy station, equipment such as smart meters, thermocouples, vibration sensors, and PLC control systems are deployed to collect real-time time-series data such as power, current, voltage, temperature, equipment switching status, and load rate.

[0058] During data collection, a combination of periodic automatic recording and event-triggered recording is used:

[0059] 1) Periodically record steady-state operating parameters at a fixed sampling frequency (e.g., 1Hz for power data, 0.1Hz for temperature data);

[0060] 2) Event triggering records will start high-frequency sampling (such as current data at 1kHz) under special operating conditions such as equipment start-up and shutdown, and load change, to ensure complete capture of transient processes.

[0061] Data cleaning removes outliers caused by sensor failures;

[0062] Energy efficiency level labels are applied to the cleaned data; three energy efficiency level labels, namely "excellent / poor / uncertain", are applied; and all collected data is stored in a structured manner according to three dimensions: timestamp, device ID, and signal type.

[0063] The labeling process is based on the preset threshold range of key performance indicators and combined with the equipment operation logic for comprehensive judgment. One implementation method is that when all indicators are simultaneously within the preset reasonable operating range, they are labeled as "excellent". If one or more key indicators continuously or seriously deviate from the preset reasonable operating range, they are labeled as "poor". If there are data conflicts or partial missing data, they are labeled as "uncertain".

[0064] S2. Perform spatiotemporal alignment processing on the original data to generate a spatiotemporal aligned feature tensor;

[0065] The multi-source sensor data from all-electric power stations includes power, temperature, and equipment switching status, which suffer from differences in sampling frequency and asynchronous timestamps. Conventional linear interpolation ignores equipment response delays, leading to causal distortion between features, while fixed-window aggregation loses high-frequency details.

[0066] This invention employs a dynamic time warping strategy to couple the physical constraints of device operation and construct a spatiotemporal alignment matrix. Specifically, it includes: for each type of equipment in the all-electric power station, calculating the minimum dynamic time warping path distance between its original time-series data and a benchmark template generated through historical data clustering; dynamically generating weighting coefficients based on this distance value and the equipment influence attenuation factor; applying the dynamic time warping algorithm to the original time-series data of each type of equipment to adjust its length to the standard time length defined by the benchmark template; and weighting and fusing the aligned results of all types of equipment based on the weighting coefficients to generate a spatiotemporal alignment feature tensor of uniform length. The specific steps are as follows:

[0067] 1) Calculate the weighting coefficient for each type of equipment.

[0068] For each type of device, the minimum dynamic time warp path distance between its original time-series data and the baseline template generated by clustering historical data is calculated. Based on this distance value, combined with the device influence attenuation factor, a weighting coefficient is dynamically generated to quantify the similarity between the device and the typical operating mode, expressed as:

[0069]

[0070] In the formula, For the first The weighting coefficients of different types of equipment reflect their similarity to typical operating modes; The device impact attenuation factor, with a default value of 0.5, controls the weight attenuation intensity of devices with high differences; For the first The minimum dynamic time-warped path distance for the type of device, i.e., for the first... Minimum dynamic time-warped path distance for all i-th samples in the class of devices Take the minimum value, where, For the first The first type of equipment The original time-series data in each sample, such as power curves; For the first The baseline template for this type of equipment is generated through historical data clustering, representing typical operating modes; This represents the total number of device types.

[0071] 2) Align all categories of device data to a standard time length.

[0072] For the raw time-series data of each type of device, a dynamic time warping algorithm is applied to adjust its length to the standard time length defined by the baseline template. Then, based on the weighted coefficients, the results of all devices after alignment are weighted and fused to generate a spatiotemporal alignment feature tensor of uniform length. This eliminates the differences in sampling frequencies and timestamp asynchrony between different devices, while preserving the temporal correlation features under physical constraints, represented as:

[0073]

[0074] In the formula, For the first The spatiotemporal alignment feature tensor after each sample alignment represents the fusion of information from all devices into a comprehensive time series. The dynamic time warping algorithm transforms time series of unequal lengths. Align to baseline template length ; To standardize the time length, a reference template is used. The time dimension is determined.

[0075] It should be noted that the present invention is achieved through... Dynamic weighting based on device operational similarity preserves cross-device temporal correlations under physical constraints, avoiding causal distortion caused by interpolation.

[0076] It should also be noted that, The output is of length . Vectors, which will contain time sequences of unequal length Aligned and scaled to the baseline template using dynamic time warping. The same length T, therefore, for the first Type of equipment, The output is a one-dimensional vector of length T.

[0077] S3. Perform multi-scale one-dimensional convolution operation on the spatiotemporal aligned feature tensor based on dynamic convolution kernel, and fuse the convolution results of different scales to generate a multi-scale feature tensor.

[0078] In sensor data from all-electric power stations, there is a scale mismatch between high-frequency features such as current / voltage and low-frequency features such as temperature. Directly stitching these features together will cause the model to be biased towards high-frequency characteristics. Conventional wavelet transforms, which use a fixed preset scale, cannot adapt to fluctuations in operating conditions such as sudden changes in the power station's load.

[0079] This invention employs a condition-aware convolutional kernel group to dynamically extract multi-scale features, thus solving the feature scale mismatch problem. The specific steps are as follows:

[0080] 1) Generate adaptive multi-scale dynamic convolution kernels

[0081] The operating condition statistics of the current sample data are obtained and their mean values ​​are calculated to obtain the mean vector of operating condition statistics. This mean vector is then input into a multilayer perceptron, which calculates and outputs a convolutional kernel weight scaling factor. The learnable basic convolutional kernel is multiplied by this scaling factor to obtain a dynamic convolutional kernel. This allows the kernel size to be dynamically adjusted according to the operating conditions, adapting to fluctuations such as sudden load changes. Specifically, this can be expressed as follows:

[0082]

[0083]

[0084] In the formula, For the first Dynamic convolution kernels at different scales, dimensions and same; For the first Scale-based learnable basic convolutional kernels; Let be the mean vector of operating condition statistics for the i-th sample, including statistics such as power and load rate; The working condition adaptation function outputs the convolution kernel weight scaling factor; For use in the first Multilayer perceptron at scale.

[0085] In one embodiment, operating condition statistics include power, load rate, current, voltage, temperature, as well as device switching state-derived quantities such as operating time percentage, downtime percentage, and fault market percentage, and vibration data-derived quantities such as average vibration amplitude, vibration effective value, and vibration variance.

[0086] 2) Integrate multi-scale features from heterogeneous sensors to generate multi-scale feature tensors.

[0087] A dynamic convolution kernel is applied to perform multi-scale one-dimensional convolution operations on the spatiotemporally aligned tensor. Features are extracted using the ReLU activation function. Then, the convolution results at different scales are concatenated along the channel dimension to generate a multi-scale feature tensor. This solves the scale mismatch problem between high-frequency features such as current / voltage and low-frequency features such as temperature, and is represented as follows:

[0088]

[0089] In the formula, Let be the multi-scale feature tensor of the i-th sample; This is the channel concatenation operator, concatenating along the first dimension. Features at each scale; For the first One-dimensional convolution operation at scale; For the first The spatiotemporal alignment feature tensor after each sample alignment represents the fusion of information from all devices into a comprehensive time series. It is the ReLU activation function; This represents the total number of scales, by default. .

[0090] It should be noted that this invention dynamically modulates the convolution kernel based on operating conditions. This addresses the sensitivity of fixed-scale convolution to sudden load changes and enables adaptive matching of feature scales.

[0091] S4. Construct an energy efficiency evaluation model for the operation of the all-electric energy station. The model is trained by training multi-scale feature tensor data.

[0092] Specifically, it includes the following steps:

[0093] S401. Define the network structure of the energy efficiency evaluation model for the operation of all-electric energy stations;

[0094] Construct a hierarchical network architecture consisting of an input layer, a feature processing layer, and a decision layer;

[0095] The input layer receives a multi-scale feature tensor, and a feature selector network is used to achieve initial dimensionality reduction, outputting a low-dimensional feature vector.

[0096] The feature processing layer includes a mutual information feature compression module and a Bayesian attention module, which are used to calculate the final uncertainty-aware feature representation from the low-dimensional feature vector. The mutual information feature compression module filters high-value features through a fully connected layer and a mutual information estimator, while the Bayesian attention module generates Gaussian distributed weights based on the interaction between the work condition query vector and the feature vector.

[0097] The decision-making layer, based on the final uncertainty perception feature representation, combines evidence theory to generate the operational energy efficiency evaluation category probability and classification confidence of the current input sample, and uses negative log-likelihood to calculate the classification loss. The classification loss function is calculated by the difference between the Dirichlet distribution and the uniform distribution with KL divergence penalty parameter as the evidence value.

[0098] Preferably, a joint optimization layer is set at the end of the network, which integrates classification loss, mutual information loss and hierarchical adaptive regularization term, and end-to-end training is achieved through a Bayesian optimizer; residual connections are used between layers to avoid gradient vanishing, and the output dimension is strictly matched with the number of energy efficiency classification categories.

[0099] S402. The input layer achieves initial dimensionality reduction through a feature selector network and outputs a low-dimensional feature vector.

[0100] The multi-scale feature tensor is input into a feature selector network consisting of two fully connected layers, which maps it to obtain a low-dimensional feature vector. This initially compresses the feature dimension, providing input for subsequent mutual information constraints, as shown below:

[0101]

[0102] In the formula, Let be the low-dimensional feature vector of the i-th sample after filtering. It is a feature selector network, consisting of two fully connected layers. This represents the learnable parameters of the selector network. Let be the multi-scale feature tensor of the i-th sample.

[0103] S403, The mutual information feature compression module performs a feature selection mechanism based on mutual information;

[0104] The sensor data of the all-electric power station contains a large number of redundant features, such as the strong correlation between ambient temperature and cooling power, which leads to information overlap and overfitting risks; conventional linear dimensionality reduction methods such as principal component analysis will destroy the physical meaning of the features, which is not conducive to subsequent attribution analysis of operating energy efficiency.

[0105] This invention constructs a feature-target mutual information bottleneck, compressing redundant features and retaining high-information features through mutual information constraints, specifically as follows:

[0106] A mutual information estimator is constructed to calculate the mutual information between the multi-scale feature tensor and the target energy efficiency level label, as well as the mutual information between the multi-scale feature tensor and the low-dimensional feature vector. Then, the two mutual information values ​​are balanced by a compression intensity coefficient to form a loss function, thereby compressing redundant features while retaining high-information features, expressed as:

[0107]

[0108] In the formula, The mutual information loss function; For mutual information estimators; Let be the multi-scale feature tensor of the i-th sample; Let i be the target energy efficiency rating label for the i-th sample; Let be the low-dimensional feature vector of the i-th sample after filtering. This is the compressibility strength coefficient, with a default value of 0.8, which controls the degree of feature compression.

[0109] The mutual information estimator is defined as follows: Equivalent value A is the first input to the mutual information estimator, and B is the second input. This is a discriminator neural network used to calculate the similarity between sample pairs; This is the softplus function; The expectation of the joint distribution; It represents the expectation of the product of marginal distributions.

[0110] It should be noted that this invention maximizes... Minimize at the same time This enables the effective preservation of high-information features and the compression of redundant features.

[0111] S404, The Bayesian attention module implements the Bayesian attention mechanism;

[0112] The importance of features changes dynamically under different operating modes. Conventional point estimation of attention weights cannot quantify decision uncertainty, leading to overfitting of weights during high-noise periods.

[0113] This invention uses a probability distribution for attention weights and quantifies the confidence level of feature importance through variance. The specific steps are as follows:

[0114] 1) Generate a working condition query vector based on the mean vector of the working condition statistics, and interact with the filtered low-dimensional feature vector. Then, calculate the mean of attention weights through the mean projection matrix, and calculate the variance of attention weights through the variance projection matrix and the ReLU activation function. This dynamically generates distribution parameters representing the importance and uncertainty of the features, expressed as:

[0115]

[0116]

[0117] In the formula, The mean of the attention weights represents the expected value of the feature importance; The standard deviation of attention weights quantifies decision uncertainty; squaring it yields the variance of attention weights. This is a working condition query vector, and ,in, It is a multilayer perceptron; Let be the vector of operating condition means for the i-th sample, such as statistics like power and load rate; for The 3D feature vector; The mean projection matrix is ​​a learnable parameter. is the variance projection matrix, which is a learnable parameter; The dimension of the feature vector; for The dimension is a positive integer; The auxiliary vector for calculating the variance is a learnable parameter; This is a vector concatenation operation; This is the ReLU activation function.

[0118] 2) To quantify the uncertainty of feature importance, attention weights are modeled as a Gaussian distribution, defined by mean and variance parameters. The mean represents the expected feature importance, and the variance quantifies the decision confidence, thus providing a probabilistic representation of feature importance, expressed as:

[0119]

[0120] In the formula, For the first The attention weights for the dimensional features follow a Gaussian distribution; Indicates a Gaussian distribution. The variance of attention weights.

[0121] 3) Feature fusion and uncertainty propagation: Attention weights are sampled from a Gaussian distribution using a reparameterization technique. Then, the sampled weights are used to perform a weighted summation of the filtered feature vectors to generate the final feature representation. This allows downstream modules to perceive the uncertainty of feature importance, expressed as:

[0122]

[0123] In the formula, This represents the uncertainty-aware feature representation of the final output. Feature vectors after filtering The 1D features; m is the filtered feature vector. Number of feature dimensions; This is a vector obtained through sampling using the reparameterization technique. ; It is a constant, such as .

[0124] It should be noted that, It can explicitly quantify attention uncertainty and avoid overfitting during periods of high noise.

[0125] S405. The decision-making level implements hierarchical evidence classification;

[0126] The energy efficiency evaluation of all-electric power stations needs to distinguish between three states: "definitely excellent / definitely poor / uncertain". Conventional softmax outputs a single probability distribution, which cannot quantify the classification confidence. Cross-entropy loss forces low-confidence samples to be misclassified, which increases the risk of misjudgment in high-noise scenarios.

[0127] This invention employs evidence theory to construct a subjective logical distribution, and uses the Dirichlet distribution to model and classify uncertainties, thereby achieving confidence-perceived energy efficiency evaluation. The specific steps are as follows:

[0128] 1) Output confidence-aware classification probabilities

[0129] Based on evidence theory, the evidence value for each category is calculated using learnable weights. Then, combined with preset basic probability quality and uncertainty, a classification probability distribution is generated through normalization. Finally, the category probability and classification confidence are output simultaneously, as follows:

[0130]

[0131] In the formula, Let be the probability that the i-th sample belongs to the c-th category; The basic probability mass of the c-th category is prior knowledge and can be set to 0 or a small constant. To support the evidence value for the c-th category, and ,in, Let be the learnable weight vector for the c-th category. for transpose; This represents the uncertainty-aware feature representation of the final output. It is a normalization factor, and , This is a measure of uncertainty. The total number of categories, such as "excellent / poor / uncertain" (3 categories).

[0132] 2) Explicitly quantifying the uncertainty of classification decisions

[0133] Aggregate the evidence values ​​of each category, combine the maximum evidence value and the category probability, calculate the uncertainty measure, and then directly obtain the confidence level of the sample classification, expressed as:

[0134]

[0135] In the formula, This is a measure of uncertainty. 0 indicates complete certainty, and 1 indicates complete uncertainty; The maximum evidence value reflects the confidence level of the dominant category; The total amount of evidence represents the sufficiency of sample information; Let be the probability that the i-th sample belongs to the c-th category. This represents the total number of categories.

[0136] 3) Design the regularization loss function

[0137] The classification loss is calculated using negative log-likelihood, and the difference between the Dirichlet distribution and the uniform distribution of the evidence value is used as the KL divergence penalty parameter to suppress overfitting to low-confidence samples, as expressed in:

[0138]

[0139] In the formula, The classification loss function; The number of samples; For the i-th sample, the true category is... Corresponding evidence value; For parameters The Dirichlet distribution; It follows a uniform Dirichlet distribution and is a prior distribution; Calculate the distribution difference for KL divergence; This is the regularization strength coefficient, with a default value of 0.2, which controls the degree of penalty for uncertainty.

[0140] It should be noted that, Value quantification quantifies decision uncertainty, thereby enabling explicit confidence modeling and supporting the determination of "uncertain" states. Furthermore, KL divergence penalizes excessive evidence, thereby achieving adaptive regularization and avoiding forced classification of low-quality data.

[0141] S406 and Bayesian regularization joint training;

[0142] Sensor noise in all-electric power stations causes local jitter in the model loss surface, making conventional optimization methods prone to suboptimal solutions. Standard weight decay cannot distinguish the importance of network parameters and lacks targeted regularization for layers with high uncertainty.

[0143] This invention constructs an adaptive regularization mechanism based on attention variance, dynamically adjusts the regularization strength of different network layers, and jointly optimizes the classification and feature selection losses. The specific steps are as follows:

[0144] 1) Achieve hierarchical regularization intensity control

[0145] Divide the L2 norm of the network parameter matrix of each network layer by the square of the uncertainty metric for that layer as a regularization term. Then, combine the classification loss and mutual information loss to form the total loss function, thereby imposing stronger regularization constraints on layers with high uncertainty, as expressed as:

[0146]

[0147] In the formula, This is the total loss function; The classification loss function; The mutual information loss function; This represents the total number of network layers. For the first Layer network parameter matrix; For the first Layer uncertainty measure.

[0148] It should be noted that, Characterizing the canonical intensity coefficient, The smaller the value, the stronger the penalty.

[0149] 2) Quantifying network hierarchical decision uncertainty

[0150] Based on the weight variance of the Bayesian attention layer, the average attention variance of the batch samples in each layer is calculated, thereby obtaining a metric representing the decision confidence of each network layer, expressed as:

[0151]

[0152] In the formula, For the first The level-average uncertainty, the smaller the value, the more uncertain the decision-making at that level; The number of feature dimensions; For the i-th sample in the i-th... Layer The variance of each attention weight.

[0153] It should be noted that, Item characterization of a single sample in The sum of attention variances of the layers Item characterization of a single sample in The average attention variance of the layer; The item characterizes the batch of samples in Expected variance of the layer.

[0154] 3) Complete the parameter optimization for noise robustness.

[0155] The first-order gradient and second-order momentum estimate of the parameters are calculated using the loss function. After bias correction, the model parameters are updated by combining a hierarchical adaptive regularization term, thus achieving stable convergence under sensor noise interference. This can be expressed as:

[0156]

[0157] In the formula, , These are the model parameter vectors for the (t+1)th and tth iterations, respectively; This is the first-order momentum estimate after bias correction; This is the second-order momentum estimate after bias correction; The learning rate is set to 0.001 by default. It is a constant, such as ;

[0158] Specifically, first-order momentum estimation Calculate as follows:

[0159]

[0160] In the formula, Let be the first moment decay rate in the t-th iteration. The first-order momentum estimate for the t-th iteration is determined as follows:

[0161]

[0162] In the formula, This is the first-order moment decay rate, with a default value of 0.9; For the first-order momentum estimate of the (t-1)th iteration, For loss function For model parameters The first gradient.

[0163] Specifically, the bias-corrected second-order momentum estimate Calculate as follows:

[0164]

[0165] In the formula, Let be the second moment decay rate in the t-th iteration. The second-order momentum estimate for the t-th iteration is determined as follows:

[0166]

[0167] In the formula, This is the second-order moment decay rate, with a default value of 0.999. The second-order momentum estimation in the (t-1)th iteration It is the element-wise square of the gradient of the loss function.

[0168] It should be noted that, Uncertainty perception regularization is used to dynamically adjust the regularization strength, with high uncertainty levels corresponding to low uncertainty levels. Furthermore, regularization terms... Implement a hierarchical adaptive mechanism to achieve differentiated processing at different levels.

[0169] S407, Iterative Training and Determining Stop Iteration Conditions;

[0170] Iterative training is performed using a mini-batch gradient descent strategy. In each iteration, batches of samples are randomly sampled from the training set, and forward propagation is performed sequentially to calculate the total loss. The network parameters are then updated using backpropagation combined with the Adam optimizer.

[0171] The training process monitors the validation set performance concurrently, and the classification accuracy and uncertainty metric on the validation set are calculated every 5 iterations. Weighted evaluation indicators, such as, .

[0172] The stopping condition adopts a dual judgment mechanism: if the validation set evaluation metric does not improve for 15 consecutive rounds, or the total number of training rounds reaches the preset upper limit (300 rounds by default), then training is terminated and the optimal model is saved; if the training loss enters a plateau period, such as the loss fluctuation range being less than 1e-4 for 10 consecutive rounds, then the early stopping mechanism is activated to prevent overfitting.

[0173] S5. After processing the real-time collected data through S1-S3, input it into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation result of the all-electric energy station.

[0174] For newly collected energy station operation data, a uniform-length spatiotemporal tensor is first generated through spatiotemporal alignment, and then heterogeneous features are fused through multi-scale convolution. This data is then input into the trained hierarchical network for forward inference.

[0175] The network output contains three types of information: energy efficiency classification probability (good / poor / uncertain), uncertainty measure κ value, and attention weight distribution of each device feature.

[0176] The evaluation phase adopts a confidence-driven decision rule: if the highest category probability is greater than the preset highest category probability threshold and κ is less than the preset minimum uncertainty measure threshold, the classification result is directly adopted; if κ is greater than or equal to the preset minimum uncertainty measure threshold or the difference between the probabilities of each category is less than the preset minimum probability difference threshold, it is judged as an "uncertain" state, triggering a manual review mechanism; in this embodiment, the preset highest category probability threshold is preferably 0.7; the preset minimum uncertainty measure threshold is preferably 0.3; and the preset minimum probability difference threshold is preferably 0.2.

[0177] This invention also claims protection for an energy efficiency evaluation system for all-electric energy stations based on the aforementioned method, comprising a data acquisition module, a spatiotemporal aligned feature tensor generation module, a multi-scale feature tensor generation module, an energy efficiency evaluation model construction module for all-electric energy stations, and an energy efficiency evaluation result calculation module for all-electric energy stations.

[0178] The data acquisition module is used to collect raw data, perform data cleaning, and label the cleaned data with energy efficiency level labels.

[0179] The spatiotemporal alignment feature tensor generation module is used to perform spatiotemporal alignment processing on the original data and generate a spatiotemporal alignment feature tensor.

[0180] The multi-scale feature tensor generation module is used to perform multi-scale one-dimensional convolution operations on the spatiotemporally aligned feature tensor based on dynamic convolution kernels, and fuse the convolution results of different scales to generate a multi-scale feature tensor.

[0181] The module for constructing an energy efficiency evaluation model for the operation of a full-electric energy station is used to construct an energy efficiency evaluation model for the operation of a full-electric energy station. By training multi-scale feature tensor data, the trained energy efficiency evaluation model for the operation of the full-electric energy station is obtained.

[0182] The module for calculating the energy efficiency evaluation results of the all-electric energy station is used to extract features from real-time collected data to generate multi-scale feature tensors, which are then input into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation results of the all-electric energy station.

[0183] Example 2:

[0184] This embodiment demonstrates the advantage of effective feature preservation in spatiotemporal alignment processing of multi-source sensor data under complex operating conditions; in one embodiment, the impact of different time-series data processing methods on the accuracy of energy efficiency classification of all-electric power stations is evaluated. (See attached...) Figure 2 As shown in the figure, the classification accuracy of different temporal alignment methods is compared. A bar chart visually compares the classification accuracy performance of traditional linear interpolation, fixed window aggregation, wavelet transform, Long Short-Term Memory (LSTM) networks, and the proposed technique. As can be seen in the figure, the purple bars corresponding to the proposed technique are significantly higher than those of other methods, and the textured bar surface simulates the fluctuation characteristics of real experimental data. The experimental results clearly demonstrate that the dynamic temporal warping strategy and multi-scale fusion mechanism proposed in this invention can effectively solve the problem of asynchronous sensor data and significantly improve classification accuracy. In particular, compared with conventional methods, this invention eliminates causal distortion caused by linear interpolation and high-frequency detail loss caused by fixed window aggregation, enabling stable recognition capabilities even under complex conditions.

[0185] Example 3:

[0186] This embodiment is used to demonstrate the anti-interference capability of the Bayesian attention mechanism and the hierarchical evidence classifier; in one embodiment, as shown in the appendix... Figure 3 The figure shows a comparison of the robustness of the methods under different noise levels. The line graph illustrates the accuracy trends of traditional linear interpolation + 3-layer fully connected neural network classification, fixed window aggregation + 3-layer fully connected neural network classification, and the proposed method as noise levels increase. All three feature curves in the figure show a downward trend, but the green curve of the proposed method remains at the top and has the gentlest slope. When the noise level on the horizontal axis increases, the curve of the proposed method only shows slight fluctuations, while the curve of the conventional method shows an accelerated downward trend. This verifies the anti-interference ability of the Bayesian attention mechanism and hierarchical evidence classifier in this invention. By explicitly quantifying decision uncertainty and dynamically adjusting feature weights, the model can maintain stable performance under sensor noise interference, solving the problem of overfitting in high-noise periods by conventional methods.

[0187] Example 4:

[0188] This embodiment is used to demonstrate the advantages of hierarchical evidence classifiers in dealing with classification uncertainty; in one embodiment, as shown in the appendix... Figure 4 As shown, the distribution of uncertainty metrics for different methods is compared, revealing the decision reliability of different methods. A violin plot combined with a box plot illustrates the distribution of the uncertainty metric κ values ​​for traditional methods (conventional Support Vector Machines), Long Short-Term Memory (LSTM) networks, and the method of this invention. The figure shows that the violin body of the method of this invention is mainly concentrated in the bottom region of the vertical axis, with a compact structure and a low upper edge, indicating that its κ value is generally at a low level. In contrast, the violin body of the traditional method is significantly shifted upwards and has a wider distribution range, with the box plot showing a large number of high outliers. Experimental results demonstrate the core advantage of the hierarchical evidence classifier of this invention: by modeling classification uncertainty through Dirichlet distribution, the model can autonomously identify low-confidence samples and trigger a manual review mechanism, avoiding the forced classification errors of conventional methods on low-quality data.

[0189] Example 5:

[0190] This embodiment is used to demonstrate the advantages of the Bayesian regularization joint training mechanism in terms of convergence stability; in one embodiment, as shown in the appendix... Figure 5 The figure shows a comparison of the training convergence speeds of different methods, used to compare the training convergence characteristics of different methods, including the loss change curves of traditional linear interpolation, Long Short-Term Memory (LSTM) networks, and the present invention as the number of training epochs increases. All three curves in the figure show a downward trend, but the green curve of the present invention initially decreases the fastest and enters a stable plateau region first in the mid-to-late stages. It is worth noting that the curves of traditional methods exhibit significant high-frequency fluctuations, while the curve of the present invention has the smallest fluctuation amplitude. This indicates that the Bayesian regularized joint training mechanism designed in this invention: constructing hierarchical adaptive regularization terms through attention variance to impose targeted constraints on high-uncertainty network layers, and employing a momentum correction optimizer, effectively overcomes the problem of local jitter in the loss surface caused by sensor noise, achieving efficient and stable convergence.

[0191] Example 6:

[0192] This embodiment is used to demonstrate the advantages of multi-scale convolutional kernels in terms of model stability; in one embodiment, as shown in the appendix... Figure 6The figure shows a comparison of the weight changes of key features over time, focusing on the stability of feature importance weights. It compares the weight fluctuations over time for two key features—voltage (high frequency) and temperature (low frequency)—using a conventional method (PCA+Softmax) and this technique (quantitative-qualitative synergy). The line graph shows that during three load abrupt change events (marked by red dashed lines), the feature weights of the conventional method exhibited drastic jumps, while the weight curve of this technique maintained a smooth transition. This stability stems from the operating condition adaptive mechanism of this technique. The multi-scale convolutional kernel dynamically modulates its size through operating condition statistics, allowing feature extraction to adapt to load fluctuations, while the probabilistic weight allocation of the Bayesian attention layer suppresses noise interference. The comparison of the two load feature curves in the figure intuitively demonstrates that it effectively solves the feature sensitivity caused by the fixed convolutional kernel in conventional methods, providing a consistent and reliable feature interpretation basis for energy station energy efficiency evaluation.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for evaluating the operational energy efficiency of a full-electric power station based on quantity-quality synergy, characterized in that, include: S1. Collect raw data and clean the data, then label the cleaned data with energy efficiency rating labels. S2. Perform spatiotemporal alignment processing on the original data to generate a spatiotemporal aligned feature tensor; For each type of equipment in the all-electric energy station, the minimum dynamic time warp path distance between its original time series data and the benchmark template generated by clustering historical data is calculated. Based on this distance value and combined with the equipment influence attenuation factor, a weight coefficient is dynamically generated. The dynamic time warping algorithm is applied to the original time series data of each type of device to adjust its length to the standard time length defined by the benchmark template. Based on the weight coefficient, the results of alignment of all types of devices are weighted and fused to generate a spatiotemporal alignment feature tensor of uniform length. S3. Perform multi-scale one-dimensional convolution operation on the spatiotemporal aligned feature tensor based on the dynamic convolution kernel, fuse the convolution results of different scales, and generate a multi-scale feature tensor; the dynamic convolution kernel is generated as follows: obtain the working condition statistics of the current sample data and calculate its mean value to obtain the working condition statistics mean vector, input the working condition statistics mean vector into the multilayer perceptron, the multilayer perceptron calculates and outputs the convolution kernel weight scaling factor, and multiplies the learnable basic convolution kernel with the convolution kernel weight scaling factor to obtain the dynamic convolution kernel; S4. Construct an energy efficiency evaluation model for the operation of the all-electric energy station. The model is trained by training multi-scale feature tensor data. S5. After processing the real-time collected data through S1-S3, input it into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation result of the all-electric energy station.

2. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 1, characterized in that, The raw data includes the power, current, voltage, temperature, equipment switching status, and load rate of each device in the all-electric power station.

3. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 2, characterized in that, The energy efficiency evaluation model for the all-electric energy station includes an input layer, a feature processing layer, and a decision layer. The input layer receives a multi-scale feature tensor and performs preliminary dimensionality reduction through a feature selector network to output a low-dimensional feature vector. The feature processing layer includes a mutual information feature compression module and a Bayesian attention module, used to calculate the final uncertainty-aware feature representation from the low-dimensional feature vector. The decision layer, based on the final uncertainty-aware feature representation and combined with evidence theory, generates the probability of the energy efficiency evaluation category and the classification confidence level corresponding to the current input sample. It then uses negative log-likelihood to calculate the classification loss and calculates the classification loss function by using the difference between the Dirichlet distribution and the uniform distribution with KL divergence penalty parameter as the evidence value.

4. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 3, characterized in that, The mutual information feature compression module specifically includes: The mutual information between the multi-scale feature tensor and the target energy efficiency level label is calculated. The difference between the product of the mutual information between the multi-scale feature tensor and the low-dimensional feature vector and the compression strength coefficient is subtracted as the mutual information loss function, which is used to compress redundant features while retaining high-information features.

5. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 4, characterized in that, The Bayesian attention module specifically includes: A working condition query vector is generated based on the mean vector of the working condition statistics, and interacts with the low-dimensional feature vector to calculate the mean of the attention weight. The variance of attention weights is calculated using the variance projection matrix and the ReLU activation function; the variance projection matrix is ​​a learnable parameter. The attention weights are modeled as a Gaussian distribution, which is defined by the mean and variance of the attention weights. Attention weights are sampled from a Gaussian distribution using a reparameterization technique. The sampled attention weights are then used to perform a weighted summation on the low-dimensional feature vector to generate the final uncertainty-aware feature representation.

6. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 5, characterized in that, In the energy efficiency evaluation model for the operation of the all-electric energy station, the L2 norm of the network parameter matrix of each network layer is calculated and divided by the square of the uncertainty metric of that layer as a regularization term; the regularization term is added to the mutual information loss function and the classification loss function to obtain the total loss function.

7. The method for evaluating the operational energy efficiency of a full-electric energy station according to claim 6, characterized in that, In S5, when the highest category probability of the energy efficiency evaluation result of the all-electric energy station is greater than the preset highest category probability threshold and the uncertainty metric value is less than the preset minimum uncertainty metric threshold, the classification result is directly adopted. When the uncertainty metric is greater than or equal to the preset minimum uncertainty metric threshold, or when the probability difference between each category is less than the preset minimum probability difference threshold, the manual review mechanism is triggered.

8. A system for evaluating the operational energy efficiency of a full-electric energy station based on the method of any one of claims 1-7, comprising a data acquisition module, a spatiotemporally aligned feature tensor generation module, a multi-scale feature tensor generation module, a full-electric energy station operational energy efficiency evaluation model construction module, and a full-electric energy station operational energy efficiency evaluation result calculation module, characterized in that: The data acquisition module is used to collect raw data, perform data cleaning, and label the cleaned data with energy efficiency level labels. The spatiotemporal alignment feature tensor generation module is used to perform spatiotemporal alignment processing on the original data and generate a spatiotemporal alignment feature tensor. The multi-scale feature tensor generation module is used to perform multi-scale one-dimensional convolution operations on the spatiotemporally aligned feature tensor based on dynamic convolution kernels, and fuse the convolution results of different scales to generate a multi-scale feature tensor. The module for constructing an energy efficiency evaluation model for the operation of a full-electric energy station is used to construct an energy efficiency evaluation model for the operation of a full-electric energy station. By training multi-scale feature tensor data, the trained energy efficiency evaluation model for the operation of the full-electric energy station is obtained. The module for calculating the energy efficiency evaluation results of the all-electric energy station is used to extract features from real-time collected data to generate multi-scale feature tensors, which are then input into the energy efficiency evaluation model of the all-electric energy station to obtain the energy efficiency evaluation results of the all-electric energy station.

Images