A Method for Inferring Aging Trends of Energy Equipment Based on Deep Prediction Models

By using the PatchTST model and the stage-controlled diffusion prediction model, the problems of insufficient generalization ability and noise sensitivity of deep prediction models in predicting the aging trend of energy equipment are solved, and high-precision and stable aging trend prediction is achieved, which is applicable to battery energy storage systems and power battery monitoring.

CN123069075APending Publication Date: 2026-11-10SHENHUA HOLLYSYS INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611292072.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing deep prediction models have insufficient generalization ability and high noise sensitivity in predicting the aging trend of energy equipment. They are difficult to capture the stage characteristics of the aging process, resulting in unstable prediction results and failing to meet long-term operation requirements.

Method used

The PatchTST model is used to extract degradation features. Combined with time series modeling, Transformer structure and stage classification mechanism, a stage-controlled diffusion prediction model is designed. Sequence prediction is performed by conditional diffusion generation method and WaveNet denoising network to achieve multi-path aging trend prediction.

Benefits of technology

It improves prediction accuracy and stability, enhances the model's adaptability to different aging stages, has good versatility and engineering adaptability, can accurately depict the aging trend of equipment, and provides reliable support for equipment life management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069075A_ABST
    Figure CN123069075A_ABST
Patent Text Reader

Abstract

This invention discloses a method for inferring the aging trend of energy equipment based on a deep prediction model, belonging to the field of energy equipment condition assessment technology. The method includes the following steps: Step 1: Acquire historical operating data of the energy equipment; Step 2: Input the time segment sequence into the PatchTST model; Step 3: Input the degradation feature vector into a pre-trained decay stage classifier, and input the degradation feature vector and stage control signal into the stage control diffusion prediction model; Step 4: Update the trainable parameters of the stage control diffusion prediction model; Step 5: In the inference stage, call the trainable parameters of the stage control diffusion prediction model to obtain the prediction sequence; Step 6: Develop the aging trend curve within the target time window. This invention combines deep time series modeling technology, diffusion probability prediction mechanism, and WaveNet denoising network structure to achieve high-precision, staged, and highly interpretable prediction and inference of the aging trend of energy equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of energy equipment condition assessment technology, and in particular to a method for inferring the aging trend of energy equipment based on a deep prediction model. Background Technology

[0002] Currently, with the widespread application of power batteries and energy storage batteries in new energy vehicles, grid energy storage, and portable devices, the issues of health status assessment and lifespan prediction during their operation are receiving increasing attention. Energy equipment undergoes performance degradation during long-term use, manifested as capacity reduction and increased internal resistance. Failure to accurately assess its aging trend may lead to reduced operating efficiency and even safety hazards. Therefore, accurate modeling and trend prediction of the aging process of energy equipment has become one of the key technologies in equipment operation and maintenance management.

[0003] In existing technologies, methods based on physical mechanism modeling are commonly used to estimate the remaining life or state of health (SOH) of batteries. For example, electrochemical models or equivalent circuit models are used to describe the internal aging behavior of batteries. However, these methods are highly dependent on equipment operating parameters, model parameters are difficult to obtain, and accuracy is low under complex operating conditions, making them unsuitable for diverse practical applications. In recent years, machine learning and deep learning-based methods have begun to be applied to predict battery aging status, especially by using multi-dimensional data such as cycle data, temperature, and voltage as input into the model for regression analysis to predict SOH or capacity retention. However, these methods mostly rely on a black-box regression framework, making it difficult to capture the stage-specific characteristics of the aging process. The model predictions lack interpretability, and the accuracy of sequence predictions fluctuates significantly.

[0004] Existing deep prediction models generally suffer from insufficient generalization ability and high sensitivity to noise when handling sequence prediction tasks, especially in the later stages of aging, where prediction errors tend to accumulate rapidly, affecting the stability of the trend curve. Furthermore, most models fail to effectively integrate aging stage information, ignoring the differences in degradation characteristics of energy equipment at different aging stages, resulting in poor stability and continuity of prediction results, making it difficult to meet the practical needs of long-term operational prediction.

[0005] Therefore, how to provide a method for inferring the aging trend of energy equipment based on a deep prediction model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for inferring the aging trend of energy equipment based on a deep prediction model. This invention fully utilizes time series modeling, a Transformer-based time feature extraction algorithm, a stage classification mechanism, and a conditional diffusion generation model. It details the entire process of extracting degradation features using the PatchTST model and combining it with decay stage control to achieve multi-path aging trend prediction. This method possesses advantages such as high prediction accuracy, strong stage adaptability, good scalability, and strong trend interpretability.

[0007] A method for inferring the aging trend of energy equipment based on a deep prediction model according to an embodiment of the present invention includes the following steps: Step 1: Obtain historical operating data of energy equipment, align the historical operating data according to time sequence, and construct a multidimensional time series dataset; Step 2: Input the multidimensional time series dataset into the PatchTST model to obtain the corresponding high-dimensional feature representation sequence. Perform aggregation operation on the high-dimensional feature representation sequence to obtain the degenerate feature vector. Step 3: Input the degradation feature vector into the decay stage classifier, encode the stage label into the stage control signal, and input the degradation feature vector and the stage control signal into the feature input module and the stage control encoding module of the stage control diffusion prediction model to construct the diffusion prediction condition vector; The stage-controlled diffusion prediction model specifically includes a feature input module, a stage-controlled coding module, a noise injection module, a denoising prediction network, a back-diffusion inference module, and a sequence output module. Step 4: During the training phase, the target sequence is associated with the diffusion prediction condition vector, the target sequence is input into the noise injection module, and a noisy target sequence is obtained based on the stage control signal. The noisy target sequence and the diffusion prediction condition vector are input into the denoising prediction network to update the trainable parameters of the stage control diffusion prediction model. Step 5: In the inference phase, the trainable parameters of the diffusion prediction model are called, and the randomly initialized noise sequence and the diffusion prediction condition vector are input into the back diffusion inference module to obtain the prediction sequence; Step 6: Input the predicted sequence into the sequence output module, which will then generate the aging trend curve of the energy equipment within the target time window in chronological order.

[0008] Optionally, the energy device is an energy storage battery.

[0009] Optionally, the historical operating data includes voltage data, current data, temperature data, charge / discharge capacity data, internal resistance data, charge / discharge cycle count, and time information.

[0010] Optionally, the step of aligning the historical running data according to time order specifically involves: Historical operational data is filtered to remove data records with missing timestamps, correct data records with abnormal timestamps, and retain data records with complete timestamps. The retained historical operating data is sorted according to time stamps, so that the various types of historical operating data of the same energy equipment are arranged in chronological order on the timeline. The sorted historical running data is resampled according to the preset time sampling interval. The missing data between adjacent time sampling points is interpolated, and the data of repeated time sampling points are merged to obtain a historical running data sequence aligned under the time sampling interval. At each time sampling point, the voltage data, current data, temperature data, charging capacity data, discharging capacity data, internal resistance data, and charge / discharge cycle count data corresponding to the time sampling point are combined to generate a feature vector containing feature components. The feature vectors corresponding to each time sampling point are arranged in chronological order to form a multidimensional time series dataset.

[0011] Optionally, step two specifically includes: The multidimensional time series dataset is divided according to the preset time window length and time step, and the feature vectors corresponding to the continuous time sampling points within each time window are combined to form a time segment. The time segments are sorted according to the order of their time sampling points to obtain a sequence of time segments arranged in chronological order. Input the time segment sequence into the PatchTST model, perform patch embedding processing on the time segment, and obtain the corresponding patch representation sequence; The patch representation sequence is subjected to positional encoding to obtain a patch-encoded sequence with positional information. Multi-head self-attention operation and feedforward network operation are performed sequentially on the patch-encoded sequence with location information to obtain a high-dimensional feature representation sequence corresponding to the time segment; Pooling operations are performed on the high-dimensional feature representation sequence in the time dimension to generate a degradation feature vector representing the current degradation state of the energy equipment.

[0012] Optionally, step three specifically includes: Input the degraded feature vector into a pre-trained decay stage classifier; The degenerate feature vector is linearly mapped and a first intermediate feature representation is generated. The first intermediate feature representation is subjected to multi-layer feature extraction processing to generate the second intermediate feature representation; The second intermediate feature is represented as a three-stage score corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. The scores of the three stages are normalized to obtain the stage probability values ​​corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. Based on the stage probability value, the stage corresponding to the stage with the highest stage probability value is selected as the aging stage of the energy equipment, and the corresponding stage label is obtained. The stage labels include the forming decay stage, the stable decay stage, and the accelerated decay stage; The stage labels are encoded and converted into stage control signals; The degradation feature vector is input into the feature input module of the stage control diffusion prediction model. The feature input module performs dimensional transformation and feature mapping on the degradation feature vector to obtain the first feature representation. The stage control signal is input into the stage control coding module of the stage control diffusion prediction model, and the stage control coding module encodes the stage control signal to obtain the second feature representation. The first feature representation and the second feature representation are concatenated according to a preset concatenation order to generate a diffusion prediction condition vector.

[0013] Optionally, step four specifically includes: During the training phase, the target sequence of the selected samples and the corresponding diffusion prediction condition vector constitute the training sample set. The target sequence is used to represent the content retention rate and health status (SOH) of the future target time window. Each target sequence is divided according to a preset number of diffusion time steps, and the target sequence segment corresponding to each time step is determined. The target sequence segment is input into the noise injection module, which calls the stage control signal corresponding to the target sequence segment. The noise intensity segment range to which the target sequence segment belongs is determined based on the stage control signal: When the stage control signal indicates the molding decay stage, the first noise intensity range is used; When the stage control signal indicates a stable decay stage, the second noise intensity range is used; When the stage control signal indicates the accelerated decay stage, the third noise intensity range is used; A linear growth strategy is applied to the noise intensity based on the time step number of the diffusion time step, and the noise intensity corresponding to the time step is determined within the noise intensity range. Based on the determined noise intensity, a mixture of Gaussian noise, Laplace noise, and impulse noise is used to perform noise superposition processing on the target sequence segment to obtain a noisy target sequence segment. Noisy target sequence segments are combined in time step order and associated with the corresponding diffusion prediction condition vector to form a noisy training sequence. The noisy training sequence is input into the denoising prediction network, and the reconstructed sequence is output at each time step; The loss value is determined based on the difference between the reconstructed sequence and the corresponding target sequence. The gradient of the trainable parameters of the denoising prediction network is calculated based on the loss value using the backpropagation algorithm of gradient descent. The Adam optimization algorithm is used to iteratively update the parameters of the denoising prediction network based on the gradient; The process of constructing noisy training sequences, generating reconstructed sequences, and updating parameters is repeated until the denoising prediction network outputs a noiseless target sequence corresponding to the target sequence at each diffusion time step, and the trainable parameters of the stage-controlled diffusion prediction model are updated.

[0014] Optionally, the training of the denoising prediction network adopts a convolutional feature extraction mechanism based on the WaveNet waveform network structure. By performing causal convolution, dilated convolution and gated activation operations on the noisy training sequence, the reconstructed sequence at each time step is generated.

[0015] Optionally, step five specifically includes: During the inference phase, the trainable parameters of the stage-controlled diffusion prediction model updated during the training phase are called, and a random initial noise sequence with the same length as the target sequence is generated according to the preset diffusion time steps. The randomly initialized noise sequence is used as the intermediate noise sequence in the last diffusion time step of the stage-controlled diffusion prediction model. The intermediate noise sequence and the diffusion prediction condition vector are input into the back diffusion inference module. In the back diffusion inference module, the back diffusion prediction operation is performed sequentially from the last diffusion time step to the first diffusion time step according to the time step number of the diffusion time step. After completing the back-diffusion prediction operation for all diffusion time steps, the denoised intermediate sequence corresponding to the first diffusion time step is used as the prediction sequence.

[0016] Optionally, the backdiffusion prediction operation specifically includes: At each diffusion time step, the noise intensity of the time step is determined according to the noise variation strategy consistent with the training phase and the corresponding phase control signal. The intermediate noise sequence of the current time step and the diffusion prediction condition vector are input into the denoising prediction network, which outputs the denoised intermediate sequence corresponding to the current time step and uses the denoised intermediate sequence as the intermediate noise sequence of the previous diffusion time step.

[0017] Optionally, step six specifically includes: In the sequence output module, the predicted sequence is parsed according to the order of time sampling points to obtain the predicted values ​​of capacity retention rate and SOH corresponding to each time sampling point within the target time window. At each time sampling point, the time information of the time sampling point is associated with the corresponding capacity retention rate prediction value. The association results are arranged according to the time order of the time sampling points to generate the first prediction data sequence of capacity retention rate changing with time. At each time sampling point, the time information of the time sampling point is associated with the corresponding health status SOH prediction value. The association results are arranged according to the time order of the time sampling points to generate a second prediction data sequence of health status SOH changing over time. Based on the first predicted data sequence, the predicted capacity retention rate of each time sampling point is sequentially connected on the time axis to generate the capacity retention rate aging trend curve of the energy equipment within the target time window; Based on the second predicted data sequence, the predicted health status (SOH) values ​​of each time sampling point are sequentially connected on the time axis to generate the SOH aging trend curve of the energy equipment within the target time window.

[0018] The beneficial effects of this invention are: This invention introduces a feature extraction method based on the PatchTST model, enabling efficient modeling and representation of historical operating data of energy equipment. It accurately extracts the current degradation state characteristics of the equipment, laying a solid foundation for subsequent aging trend prediction. This feature extraction method possesses strong time-series awareness and multi-dimensional feature fusion capabilities, exhibiting higher expression accuracy compared to traditional statistical methods and shallow neural networks.

[0019] This invention innovatively designs a decay stage classifier. By performing multi-layer nonlinear feature extraction and stage classification on the degradation feature vector, it achieves the distinction of different stages in the equipment aging process. Furthermore, the stage labels are encoded as control signals and introduced into the prediction model, enabling the subsequent diffusion prediction process to be carried out under a staged control mechanism. This enhances the model's adaptability to the characteristics of different degradation stages and improves the dynamic response capability of the prediction.

[0020] The proposed stage-controlled diffusion prediction model combines a conditional diffusion generation method with a WaveNet denoising network structure, enabling highly robust sequence prediction tasks under different stage conditions. By adjusting the noise injection intensity and variation strategy through stage control signals, stage-sensitive noise modeling is achieved, thereby giving the model stronger generalization ability in the inference stage and significantly improving the accuracy and stability of aging trend prediction.

[0021] This invention also achieves end-to-end optimization of model parameters by constructing a unified training and inference process, enabling the constructed model to have good engineering adaptability in actual deployment. The prediction results can accurately depict the changes in the equipment's content retention rate and state of health (SOH) within the target time window, providing reliable support for equipment lifespan management and maintenance strategies. This method has good versatility and promotion value, and can be widely applied in fields such as battery energy storage systems and power battery monitoring. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a method for inferring the aging trend of energy equipment based on a deep prediction model, as proposed in this invention. Figure 2 This is a schematic diagram of the stage-controlled diffusion prediction model structure of the energy equipment aging trend inference method based on deep prediction model proposed in this invention. Figure 3 This is a flowchart of the reverse diffusion prediction process in the inference stage of an energy equipment aging trend inference method based on a deep prediction model proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figures 1-3 A method for inferring the aging trend of energy equipment based on a deep prediction model includes the following steps: Step 1: Obtain historical operating data of energy equipment, align the historical operating data according to time sequence, and construct a multidimensional time series dataset; Step 2: Process and sort the multidimensional time series dataset to form a time segment sequence. Input the time segment sequence into the PatchTST model to obtain the corresponding high-dimensional feature representation sequence. Perform aggregation operation on the high-dimensional feature representation sequence to obtain the degradation feature vector representing the current degradation state of the energy equipment. Step 3: Input the degradation feature vector into the pre-trained decay stage classifier to classify the aging stages of the energy equipment and obtain stage labels. Encode the stage labels into stage control signals. Input the degradation feature vector and stage control signals into the feature input module and stage control encoding module of the stage control diffusion prediction model to construct the diffusion prediction condition vector. Step 4: During the training phase, the target sequence representing the content retention rate and health status (SOH) of the future target time window is associated with the diffusion prediction condition vector. The target sequence is input into the noise injection module. According to the stage control signal, the corresponding noise intensity and noise change strategy are set to obtain the noisy target sequence. The noisy target sequence and the diffusion prediction condition vector are input into the denoising prediction network to train the denoising prediction network, reconstruct the corresponding noise-free target sequence, and update the trainable parameters of the stage control diffusion prediction model. Step 5: In the inference phase, obtain the corresponding diffusion prediction condition vector, call the trainable parameters of the stage-controlled diffusion prediction model, use the randomly initialized noise sequence as the initial input sequence of the stage-controlled diffusion prediction model, input the randomly initialized noise sequence and the diffusion prediction condition vector into the back diffusion inference module, remove the noise in the randomly initialized noise sequence, and obtain the prediction sequence. Step 6: Input the predicted sequence into the sequence output module, which will then generate the aging trend curve of the energy equipment within the target time window in chronological order.

[0025] In this embodiment, the stage-controlled diffusion prediction model specifically includes a feature input module, a stage-controlled coding module, a noise injection module, a denoising prediction network, a backdiffusion inference module, and a sequence output module.

[0026] The noise injection uses a mixture of Gaussian noise, Laplace noise, and impulse noise; In this embodiment, step one specifically includes: The energy device is an energy storage battery, and its historical operating data includes voltage data, current data, temperature data, charge / discharge capacity data, internal resistance data, charge / discharge cycle count, and time information. The historical data is aligned according to chronological order, specifically as follows: Historical operational data is filtered to remove data records with missing timestamps, correct data records with abnormal timestamps, and retain data records with complete timestamps. The retained historical operating data is sorted according to time stamps, so that the various types of historical operating data of the same energy equipment are arranged in chronological order on the timeline. The sorted historical running data is resampled according to the preset time sampling interval. The missing data between adjacent time sampling points is interpolated, and the data of repeated time sampling points are merged to obtain a historical running data sequence aligned under the time sampling interval. The preset time sampling interval can be adjusted according to the monitoring frequency of the energy equipment to resample historical operating data and form a time-aligned historical operating data sequence.

[0027] At each time sampling point, the voltage data, current data, temperature data, charging capacity data, discharging capacity data, internal resistance data, and charge / discharge cycle count data corresponding to the time sampling point are combined to generate a feature vector containing feature components. The feature vectors corresponding to each time sampling point are arranged in chronological order to form a multidimensional time series dataset.

[0028] In the specific implementation process, the time sampling interval is set to 30 minutes. Voltage, current, temperature, charging capacity, discharging capacity, internal resistance, and charge / discharge cycle count are subjected to maximum-minimum normalization before forming the feature vector. Specifically, for each data category, the minimum value of the corresponding category within the 180-day collection period is subtracted, and then divided by the difference between the maximum and minimum values ​​of the corresponding category, thus limiting each feature component to the range of 0 to 1. This processing reduces the impact of different data dimensions on subsequent model inputs, enabling the multi-dimensional time series dataset to more stably reflect energy storage battery capacity decay, internal resistance changes, and temperature fluctuations, thereby improving the consistency of degradation feature extraction.

[0029] In this embodiment, step two specifically includes: The multidimensional time series dataset is divided according to the preset time window length and time step, and the feature vectors corresponding to the continuous time sampling points within each time window are combined to form a time segment. The time segments are sorted according to the order of their time sampling points to obtain a sequence of time segments arranged in chronological order. Input the time segment sequence into the PatchTST model, perform patch embedding processing on the time segment, and obtain the corresponding patch representation sequence; The patch representation sequence is subjected to positional encoding to obtain a patch-encoded sequence with positional information. Multi-head self-attention operation and feedforward network operation are performed sequentially on the patch-encoded sequence with location information to obtain a high-dimensional feature representation sequence corresponding to the time segment; Pooling operations are performed on the high-dimensional feature representation sequence in the time dimension to generate a degradation feature vector representing the current degradation state of the energy equipment.

[0030] In the specific implementation, the time window length is set to 24 hours, the time step is set to 12 hours, each time segment contains 48 time sampling points, and adjacent time segments retain 24 overlapping sampling points to ensure continuous input of the aging process into the model. The patch length is set to 8 sampling points, each time segment is divided into 6 patches, the patch embedding dimension is set to 64 dimensions, the number of multi-head self-attention heads is set to 4, and the hidden dimension of the feedforward network is set to 128 dimensions. Through the above settings, the intraday fluctuation characteristics and cross-day degradation trends of the energy storage battery can be taken into account, avoiding insufficient degradation information due to an excessively short time window, and also avoiding the smoothing of local anomalies due to an excessively long window, thereby improving the ability of the degradation feature vector to represent the current aging state.

[0031] In this embodiment, step three specifically includes: Input the degraded feature vector into a pre-trained decay stage classifier; The degenerate feature vector is linearly mapped and a first intermediate feature representation is generated. The first intermediate feature representation is subjected to multi-layer feature extraction processing to generate the second intermediate feature representation; The multi-layer feature extraction process for the first intermediate feature representation is as follows: The first intermediate feature representation is sequentially input into multiple fully connected layers. Each fully connected layer performs a linear mapping on the first intermediate feature representation. An activation function layer is set between each fully connected layer to perform a nonlinear transformation on each linear mapping result, thereby generating a second intermediate feature representation.

[0032] The second intermediate feature is represented as a three-stage score corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. The second intermediate feature representation is input into the classification output layer, which includes a fully connected layer for performing linear mapping operations. The fully connected layer maps the second intermediate feature representation from a preset dimension into three output components, which correspond to the scores of the shaping decay stage, the stable decay stage, and the accelerated decay stage, respectively.

[0033] The scores of the three stages are normalized to obtain the stage probability values ​​corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. Based on the stage probability value, the stage corresponding to the stage with the highest stage probability value is selected as the aging stage of the energy equipment, and the corresponding stage label is obtained. The stage labels include the forming decay stage, the stable decay stage, and the accelerated decay stage; The stage labels are encoded and converted into stage control signals; The degraded feature vector is input into the feature input module of the stage control diffusion prediction model. The feature input module performs dimensional transformation and feature mapping on the degraded feature vector to obtain the first feature representation. The specific processes of dimensional transformation and feature mapping are as follows: The degenerate feature vector is input into the first fully connected layer, and a linear mapping process is performed on the degenerate feature vector to convert the degenerate feature vector from the original feature dimension to the preset feature dimension; The feature vector processed by the first fully connected layer is input into the activation function layer, and a nonlinear transformation is performed on the feature vector to obtain the first feature representation.

[0034] The stage control signal is input into the stage control coding module of the stage control diffusion prediction model, and the stage control coding module encodes the stage control signal to obtain the second feature representation. The first feature representation and the second feature representation are concatenated according to a preset concatenation order to generate a diffusion prediction condition vector.

[0035] In the specific implementation, the degradation feature vector input to the decay stage classifier is set to 64 dimensions. After linear mapping, a 128-dimensional first intermediate feature representation is obtained. Multi-layer feature extraction uses two fully connected layers, with output dimensions of 128 and 64 dimensions respectively. The ReLU function is used as the activation function. The classification output layer outputs three stage scores, and three stage probability values ​​are obtained through Softmax normalization. The stage corresponding to the highest stage probability value is determined as the current aging stage. The stage label is converted into a stage control signal using one-hot encoding. The shaping decay stage is encoded as [1,0,0], the stable decay stage is encoded as [0,1,0], and the accelerated decay stage is encoded as [0,0,1]. In the stage-controlled diffusion prediction model, the degradation feature vector is mapped to a 64-dimensional first feature representation through the feature input module, and the stage control signal is mapped to a 16-dimensional second feature representation through the stage control encoding module. The two are concatenated according to their feature dimensions to generate an 80-dimensional diffusion prediction condition vector. The above settings enable the diffusion prediction process to simultaneously receive information on the degree of degradation and the aging stage, avoiding trend deviation caused by using the same prediction conditions for different decay stages and improving the stage adaptability of subsequent aging trend predictions.

[0036] In this embodiment, step four specifically includes: During the training phase, the target sequence of the selected samples and the corresponding diffusion prediction condition vector constitute the training sample set. The target sequence is used to represent the content retention rate and health status (SOH) of the future target time window. The samples are derived from historical operating data collected during the actual operation of energy equipment and cyclic decay test data collected under experimental conditions. The target sequence is constructed based on the capacity retention rate data and state of health (SOH) data actually collected by the energy equipment within the future target time window, and is used as a real sequence to represent the changes of aging indicators over time during the training phase.

[0037] Each target sequence is divided according to a preset number of diffusion time steps, and the target sequence segment corresponding to each time step is determined. The target sequence segment is input into the noise injection module, which calls the stage control signal corresponding to the target sequence segment. The noise intensity segment range to which the target sequence segment belongs is determined based on the stage control signal: When the stage control signal indicates the molding decay stage, the first noise intensity range is used; When the stage control signal indicates a stable decay stage, the second noise intensity range is used; When the stage control signal indicates the accelerated decay stage, the third noise intensity range is used; First noise intensity range , used to correspond to the molding attenuation stage; Second noise intensity range This is used to correspond to the stable decay phase; Third noise intensity range This is used to correspond to the accelerated decay phase.

[0038] A linear growth strategy is applied to the noise intensity based on the time step number of the diffusion time step, and the noise intensity corresponding to the time step is determined within the noise intensity range. Based on the determined noise intensity, a mixture of Gaussian noise, Laplace noise, and impulse noise is used to perform noise superposition processing on the target sequence segment to obtain a noisy target sequence segment. Noisy target sequence segments are combined in time step order and associated with the corresponding diffusion prediction condition vector to form a noisy training sequence. The noisy training sequence is input into the denoising prediction network, and the reconstructed sequence is output at each time step; The loss value is determined based on the difference between the reconstructed sequence and the corresponding target sequence. The gradient of the trainable parameters of the denoising prediction network is calculated based on the loss value using the backpropagation algorithm of gradient descent. The Adam optimization algorithm is used to iteratively update the parameters of the denoising prediction network based on the gradient; the learning rate of the Adam optimization algorithm is set to 0.001, the batch size is set to 32, and the number of training rounds is set to 100. The process of constructing noisy training sequences, generating reconstructed sequences, and updating parameters is repeated until the denoising prediction network outputs a noiseless target sequence corresponding to the target sequence at each diffusion time step, and the trainable parameters of the stage-controlled diffusion prediction model are updated.

[0039] The trainable parameters include the weight and bias parameters of the linear mapping layer in the feature input module, the weight and bias parameters of the coding layer in the stage control coding module, the kernel weights and biases of the causal convolutional and dilated convolutional layers in the denoising prediction network, the weight and bias parameters of the gated activation units, the weight parameters of the residual connection units and skip connection units, and the weight and bias parameters of the prediction network in the backdiffusion inference module.

[0040] In this embodiment, the training of the denoising prediction network adopts a convolutional feature extraction mechanism based on the WaveNet waveform network structure. By performing causal convolution, dilated convolution and gated activation operations on the noisy training sequence, the reconstructed sequence at each time step is generated. This invention, by introducing a stage control mechanism and a WaveNet-based denoising prediction network, achieves refined modeling and sequence reconstruction of different aging stages in diffusion prediction modeling, enhancing the perception and prediction capabilities of aging trends. Through a three-stage noise intensity adjustment and hybrid noise injection strategy, the robustness of the model in complex environments is improved; the convolutional structure of WaveNet enhances the accuracy of time series reconstruction, making the prediction results smoother and more stable. The overall method achieves high-precision fitting and long-term prediction of aging trends, effectively improving the intelligence level of energy equipment health status assessment.

[0041] In this embodiment, step five specifically includes: In the inference phase, historical operating data of the energy equipment to be evaluated is acquired and processed according to steps one and two to form a multidimensional time series dataset and degradation feature vector; The degradation feature vector is input into the pre-trained decay stage classifier according to step three to obtain the stage label and encode the stage control signal. The degradation feature vector and the stage control signal are then input into the feature input module and stage control encoding module of the stage control diffusion prediction model to generate the corresponding diffusion prediction condition vector. The trainable parameters of the stage-controlled diffusion prediction model updated during the training phase are called, and a random initialization noise sequence with the same length as the target sequence is generated according to the preset diffusion time step. The random initialization noise sequence is used as the intermediate noise sequence of the stage-controlled diffusion prediction model in the last diffusion time step. The intermediate noise sequence and the diffusion prediction condition vector are input into the back diffusion inference module. In the back diffusion inference module, the back diffusion prediction operation is performed sequentially from the last diffusion time step to the first diffusion time step according to the time step number of the diffusion time step. The backdiffusion prediction operation is specifically as follows: at each diffusion time step, the noise intensity of the time step is determined according to the noise change strategy consistent with the training phase and the corresponding phase control signal. The intermediate noise sequence of the current time step and the diffusion prediction condition vector are input into the denoising prediction network, which outputs the denoised intermediate sequence corresponding to the current time step and uses the denoised intermediate sequence as the intermediate noise sequence of the previous diffusion time step. After completing the reverse diffusion prediction operation for all diffusion time steps, the denoised intermediate sequence corresponding to the first diffusion time step is used as the prediction sequence. The prediction sequence is used to represent the prediction results of the content retention rate and health status (SOH) of the future target time window as a function of time.

[0042] The specific process of generating the denoised intermediate sequence is as follows: Input the intermediate noise sequence at the current time step into the input of the denoising prediction network; In the denoising prediction network, causal convolution, dilated convolution, and gated activation operations are performed sequentially on the input sequence. When performing causal convolution operations, the kernel weight parameters and bias parameters of the causal convolution layer are called to perform convolution processing on the input sequence. When performing dilated convolution operations, the kernel weight parameters and bias parameters of the dilated convolution layer are called to perform convolution processing on the sequence obtained by the causal convolution operation. When performing the gated activation operation, the weight parameters and bias parameters of the gated activation unit are called to perform gated processing on the sequence obtained by the dilated convolution operation; The residual connection unit is invoked to perform residual fusion on the output of each convolutional layer, and the skip connection unit is invoked to perform feature aggregation on the convolutional results, generating the denoised intermediate sequence for the corresponding time step; The denoised intermediate sequence is used as the intermediate noise sequence of the previous diffusion time step and input into the back diffusion inference module to continue the back diffusion prediction operation.

[0043] In the inference phase, this invention constructs a backdiffusion inference module. Utilizing the updated stage-controlled diffusion prediction model parameters from the training phase, and combining a preset diffusion time step with a randomly initialized noise sequence, it progressively executes backdiffusion prediction operations. Within each diffusion time step, a denoising prediction network based on the WaveNet structure sequentially performs causal convolution, dilated convolution, gated activation, residual connections, and skip connections to generate a denoised intermediate sequence, which serves as the input to the previous time step. The final output is a predicted sequence representing the capacity retention rate and the trend of SOH changes, achieving high-precision intelligent prediction of the future aging state of energy equipment.

[0044] In this embodiment, step six specifically includes: In the sequence output module, the predicted sequence is parsed according to the order of time sampling points to obtain the predicted values ​​of capacity retention rate and SOH corresponding to each time sampling point within the target time window. In the sequence output module, the predicted sequence is parsed according to the chronological order of the time sampling points, specifically as follows: The sequence output module is called to perform index parsing processing on the predicted sequence, and the vector elements corresponding to each time step in the predicted sequence are read. For each time step, the vector elements are subjected to dimension separation processing. The first dimension of the vector elements is used as the capacity retention rate prediction value, and the second dimension of the vector elements is used as the health status (SOH) prediction value. The capacity retention rate prediction values ​​are arranged in chronological order of the sampling points to form a capacity retention rate prediction data sequence. The predicted SOH values ​​for each health status are arranged in chronological order according to the sampling points to form a predicted SOH data sequence for the health status.

[0045] At each time sampling point, the time information of the time sampling point is associated with the corresponding capacity retention rate prediction value. The association results are arranged according to the time order of the time sampling points to generate the first prediction data sequence of capacity retention rate changing with time. At each time sampling point, the time information of the time sampling point is associated with the corresponding health status SOH prediction value. The association results are arranged according to the time order of the time sampling points to generate a second prediction data sequence of health status SOH changing over time. Based on the first predicted data sequence, the predicted capacity retention rate of each time sampling point is sequentially connected on the time axis to generate the capacity retention rate aging trend curve of the energy equipment within the target time window; Based on the second predicted data sequence, the predicted health status (SOH) values ​​of each time sampling point are sequentially connected on the time axis to generate the SOH aging trend curve of the energy equipment within the target time window.

[0046] Example 1: To verify the practical application effect of the present invention, it was applied to the energy storage battery equipment in operation at a new energy company. A 180-day aging trend prediction experiment was conducted, focusing on solving the problems of traditional methods, such as inability to dynamically adapt to different aging stages, low prediction accuracy, and slow response. The company's energy storage system operates in a high-temperature, multi-cycle environment, with battery modules undergoing multiple charge-discharge switching daily. The ambient temperature fluctuates between 28°C and 46°C. Some batteries exhibited accelerated capacity retention decline, increased internal resistance, and discontinuous SOH changes during high-load operation. Traditional linear regression methods can only extrapolate the overall capacity decline trend based on historical data, making it difficult to identify the differences between forming degradation, stable degradation, and accelerated degradation in a timely manner, resulting in a delayed warning of aging inflection points.

[0047] Hundreds of energy storage battery modules were deployed at the experimental site. Historical operational data from 50 battery modules were collected as the basis for training and validation. This data included voltage, current, temperature, charge / discharge capacity, internal resistance, cycle count, and corresponding timestamps. The sampling interval was set to 30 minutes, covering the entire process from equipment installation and operation to testing. During data preprocessing, historical data underwent time alignment, anomaly removal, and interpolation repair to construct a complete multidimensional time series dataset. The time series was then divided into multiple continuous time segments, each with a 24-hour window and a 12-hour step size. These segments were input into the PatchTST model to generate a high-dimensional feature representation sequence. Features were aggregated through time pooling to obtain a degradation feature vector representing the degree of equipment degradation.

[0048] In the model application process, a pre-trained degradation stage classifier classifies the degradation feature vector into stages, labeling the current battery state as forming degradation, stable degradation, or accelerated degradation, and generating corresponding stage control signals. During the training phase, the actual measured capacity retention rate and SOH are used to construct the target sequence. Different intensities of mixed noise are superimposed on the stage control signals, with the mixing ratio of Gaussian noise, Laplace noise, and impulse noise set to 6:3:1. Gaussian noise is used to simulate normal sensor sampling fluctuations, Laplace noise to simulate non-stationary disturbances caused by charge / discharge switching and load fluctuations, and impulse noise to simulate communication jitter or instantaneous sampling spikes. This ratio setting enhances the denoising prediction network's adaptability to complex operational disturbances while maintaining the continuity of the aging trend. The WaveNet denoising prediction network is trained using diffusion prediction conditional vectors and noisy training sequences. The training process employs the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 epochs. The final error level on the validation set is 1.72% MAE and 2.48% RMSE.

[0049] To verify the practical effectiveness of this invention, a comparative test was conducted on the traditional linear regression method, the conventional LSTM prediction method, and the method of this invention using the same batch of energy storage battery operating data. The traditional linear regression method uses the historical capacity retention rate and state of equilibrium (SOH) trends over time as input to directly fit the aging curve for the next 7 days. The conventional LSTM prediction method uses a multi-dimensional sequence composed of voltage, current, temperature, capacity, internal resistance, and cycle number as input to output predicted values ​​for capacity retention rate and SOH for the next 7 days. The method of this invention further extracts degradation feature vectors using the PatchTST model based on the multi-dimensional time series dataset, generates stage control signals using a decay stage classifier, and completes the inference of aging trends with stage constraints through a stage-controlled diffusion prediction model. During the test, 10 battery modules that were not trained were selected for verification. For each battery module, the capacity retention rate and SOH trends for the next 7 days, totaling 336 time sampling points, were predicted.

[0050] Test results show that the traditional linear regression method has an average absolute error of 3.86% in capacity retention and 4.12% in state of equilibrium (SOH) across 10 battery groups, with a maximum error of 7.35% during accelerated degradation and an average lag of 3.6 days towards the aging inflection point. The conventional LSTM prediction method has an average absolute error of 2.74% in capacity retention and 2.91% in SOH, with a maximum error of 5.18% during accelerated degradation and an average lag of 2.1 days towards the aging inflection point. The method of this invention has an average absolute error of 1.38% in capacity retention and 1.56% in SOH, with the maximum error during accelerated degradation controlled at 2.62%, and provides an average warning of the aging inflection point 1.4 days earlier. Specifically, for the energy storage battery module numbered B17, the actual capacity retention on day 1 was 84.9%, decreasing to 82.8% on day 7. The traditional linear regression predicted a capacity retention of 84.1% on day 7, the conventional LSTM predicted 83.6%, and this invention predicted 83.1%. The actual SOH decreased from 84.2% to 82.1%, while the traditional linear regression prediction was 83.7%, the conventional LSTM prediction was 83.0%, and the prediction of this invention was 82.4%.

[0051] The experiments above show that traditional methods, once they enter the accelerated degradation stage, suffer from significantly increased errors due to the inability to adjust prediction strategies according to the aging stage, and also exhibit a lag in response to the aging inflection point. In contrast, the method of this invention utilizes the PatchTST model to extract multidimensional degradation features and uses the degradation stage classification results as stage control signals to input into the stage control diffusion prediction model. This allows noise injection and back diffusion inference to adjust noise intensity according to the aging stage, and integrates the denoising and reconstruction process with the diffusion prediction condition vector. This enables more stable output of the capacity retention rate and state of health (SOH) trend curves within the future target time window, significantly improving the accuracy, robustness, and early warning capability of energy storage battery aging trend prediction.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for inferring the aging trend of energy equipment based on a deep prediction model, characterized in that, Includes the following steps: Step 1: Obtain historical operating data of energy equipment, align the historical operating data according to time sequence, and construct a multidimensional time series dataset; Step 2: Input the multidimensional time series dataset into the PatchTST model to obtain the corresponding high-dimensional feature representation sequence. Perform aggregation operation on the high-dimensional feature representation sequence to obtain the degenerate feature vector. Step 3: Input the degradation feature vector into the decay stage classifier, encode the stage label into the stage control signal, and input the degradation feature vector and the stage control signal into the feature input module and the stage control encoding module of the stage control diffusion prediction model to construct the diffusion prediction condition vector; The stage-controlled diffusion prediction model specifically includes a feature input module, a stage-controlled coding module, a noise injection module, a denoising prediction network, a back-diffusion inference module, and a sequence output module. Step 4: During the training phase, the target sequence is associated with the diffusion prediction condition vector, the target sequence is input into the noise injection module, and a noisy target sequence is obtained based on the stage control signal. The noisy target sequence and the diffusion prediction condition vector are input into the denoising prediction network to update the trainable parameters of the stage control diffusion prediction model. Step 5: In the inference phase, the trainable parameters of the diffusion prediction model are called, and the randomly initialized noise sequence and the diffusion prediction condition vector are input into the back diffusion inference module to obtain the prediction sequence; Step 6: Input the predicted sequence into the sequence output module, which will then generate the aging trend curve of the energy equipment within the target time window in chronological order.

2. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, The energy device is an energy storage battery.

3. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, The historical operating data includes voltage data, current data, temperature data, charge / discharge capacity data, internal resistance data, charge / discharge cycle count, and time information.

4. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, The process of aligning historical operational data according to chronological order to construct a multidimensional time series dataset is as follows: Historical operational data is filtered to remove data records with missing timestamps, correct data records with abnormal timestamps, and retain data records with complete timestamps. The retained historical operating data is sorted according to time stamps, so that the various types of historical operating data of the same energy equipment are arranged in chronological order on the timeline. The sorted historical running data is resampled according to the preset time sampling interval. The missing data between adjacent time sampling points is interpolated, and the data of repeated time sampling points are merged to obtain a historical running data sequence aligned under the time sampling interval. At each time sampling point, the voltage data, current data, temperature data, charging capacity data, discharging capacity data, internal resistance data, and charge / discharge cycle count data corresponding to the time sampling point are combined to generate a feature vector containing feature components. The feature vectors corresponding to each time sampling point are arranged in chronological order to form a multidimensional time series dataset.

5. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, Step two specifically involves: The multidimensional time series dataset is divided according to the preset time window length and time step, and the feature vectors corresponding to the continuous time sampling points within each time window are combined to form a time segment. The time segments are sorted according to the order of their time sampling points to obtain a sequence of time segments arranged in chronological order. Input the time segment sequence into the PatchTST model, perform patch embedding processing on the time segment, and obtain the corresponding patch representation sequence; The patch representation sequence is subjected to positional encoding to obtain a patch-encoded sequence with positional information. Multi-head self-attention operation and feedforward network operation are performed sequentially on the patch-encoded sequence with location information to obtain a high-dimensional feature representation sequence corresponding to the time segment; Pooling operations are performed on the high-dimensional feature representation sequence in the time dimension to generate a degradation feature vector representing the current degradation state of the energy equipment.

6. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, Step three specifically involves: Input the degraded feature vector into a pre-trained decay stage classifier; The degenerate feature vector is linearly mapped and a first intermediate feature representation is generated. The first intermediate feature representation is subjected to multi-layer feature extraction processing to generate the second intermediate feature representation; The second intermediate feature is represented as a three-stage score corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. The scores of the three stages are normalized to obtain the stage probability values ​​corresponding to the forming decay stage, the stable decay stage, and the accelerated decay stage, respectively. Based on the stage probability value, the stage corresponding to the stage with the highest stage probability value is selected as the aging stage of the energy equipment, and the corresponding stage label is obtained. The stage labels include the forming decay stage, the stable decay stage, and the accelerated decay stage; The stage labels are encoded and converted into stage control signals; The degradation feature vector is input into the feature input module of the stage control diffusion prediction model. The feature input module performs dimensional transformation and feature mapping on the degradation feature vector to obtain the first feature representation. The stage control signal is input into the stage control coding module of the stage control diffusion prediction model, and the stage control coding module encodes the stage control signal to obtain the second feature representation. The first feature representation and the second feature representation are concatenated according to a preset concatenation order to generate a diffusion prediction condition vector.

7. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, Step four specifically involves: During the training phase, the target sequence of the selected samples and the corresponding diffusion prediction condition vector constitute the training sample set. The target sequence is used to represent the content retention rate and health status (SOH) of the future target time window. Each target sequence is divided according to a preset number of diffusion time steps, and the target sequence segment corresponding to each time step is determined. The target sequence segment is input into the noise injection module, which calls the stage control signal corresponding to the target sequence segment. The noise intensity segment range to which the target sequence segment belongs is determined based on the stage control signal; A linear growth strategy is applied to the noise intensity based on the time step number of the diffusion time step, and the noise intensity corresponding to the time step is determined within the noise intensity range. Based on the determined noise intensity, a mixture of Gaussian noise, Laplace noise, and impulse noise is used to perform noise superposition processing on the target sequence segment to obtain a noisy target sequence segment. Noisy target sequence segments are combined in time step order and associated with the corresponding diffusion prediction condition vector to form a noisy training sequence. The noisy training sequence is input into the denoising prediction network. The training of the denoising prediction network adopts a convolutional feature extraction mechanism based on the WaveNet waveform network structure. By performing causal convolution, dilated convolution and gated activation operations on the noisy training sequence, the reconstructed sequence at each time step is generated. The loss value is determined based on the difference between the reconstructed sequence and the corresponding target sequence. The gradient of the trainable parameters of the denoising prediction network is calculated based on the loss value using the backpropagation algorithm of gradient descent. The Adam optimization algorithm is used to iteratively update the parameters of the denoising prediction network based on the gradient; The process of constructing noisy training sequences, generating reconstructed sequences, and updating parameters is repeated until the denoising prediction network outputs a noise-free target sequence corresponding to the target sequence at each diffusion time step, and the trainable parameters of the stage-controlled diffusion prediction model are updated.

8. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, Step five specifically involves: During the inference phase, the trainable parameters of the stage-controlled diffusion prediction model updated during the training phase are called, and a random initial noise sequence with the same length as the target sequence is generated according to the preset diffusion time steps. The randomly initialized noise sequence is used as the intermediate noise sequence in the last diffusion time step of the stage-controlled diffusion prediction model. The intermediate noise sequence and the diffusion prediction condition vector are input into the back diffusion inference module. In the back diffusion inference module, the back diffusion prediction operation is performed sequentially from the last diffusion time step to the first diffusion time step according to the time step number of the diffusion time step. After completing the back-diffusion prediction operation for all diffusion time steps, the denoised intermediate sequence corresponding to the first diffusion time step is used as the prediction sequence.

9. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 8, characterized in that, The reverse diffusion prediction operation is specifically as follows: At each diffusion time step, the noise intensity of the time step is determined according to the noise variation strategy consistent with the training phase and the corresponding phase control signal. The intermediate noise sequence of the current time step and the diffusion prediction condition vector are input into the denoising prediction network, which outputs the denoised intermediate sequence corresponding to the time step and uses the denoised intermediate sequence as the intermediate noise sequence of the previous diffusion time step.

10. The method for inferring the aging trend of energy equipment based on a deep prediction model according to claim 1, characterized in that, Step six specifically involves: In the sequence output module, the predicted sequence is parsed according to the order of time sampling points to obtain the predicted values ​​of capacity retention rate and SOH corresponding to each time sampling point within the target time window. At each time sampling point, the time information of the time sampling point is associated with the corresponding capacity retention rate prediction value. The association results are arranged according to the time order of the time sampling points to generate the first prediction data sequence of capacity retention rate changing with time. At each time sampling point, the time information of the time sampling point is associated with the corresponding health status SOH prediction value. The association results are arranged according to the time order of the time sampling points to generate a second prediction data sequence of health status SOH changing over time. Based on the first predicted data sequence, the predicted capacity retention rate of each time sampling point is sequentially connected on the time axis to generate the capacity retention rate aging trend curve of the energy equipment within the target time window; Based on the second predicted data sequence, the predicted health status (SOH) values ​​of each time sampling point are sequentially connected on the time axis to generate the SOH aging trend curve of the energy equipment within the target time window.