Classification processing system for energy storage data

Abstract

The invention discloses a hierarchical processing system for energy storage data, and the system comprises an optimization feature extraction module which collects initial energy storage data, builds a unified time sequence on the time dimension of the initial energy storage data through a specified time interval # imgabs0 #, carries out the alignment sampling, introduces the initial energy storage data after the alignment sampling into an auto-encoder, and carries out the noise reduction processing, performing feature extraction on the data subjected to noise reduction processing to obtain optimized energy storage feature data; the dynamic grading adjustment module is used for standardizing the optimized energy storage characteristic data, calculating the information entropy of the standardized characteristic to obtain an initial weight, introducing a Q reinforcement learning algorithm to carry out online optimization on the weight, designing a third-order grading model based on the optimized initial weight, and carrying out dynamic grading adjustment on the optimal energy storage characteristic data; and performing data classification on different optimization energy storage characteristic data combinations and dynamic weights according to a third-order classification model, dividing the data into an emergency level, an important level and a common level based on the third-order classification model, and providing targeted processing and management strategies for different levels of data.

Description

Technical Field

[0001] The present invention relates to the technical field of data hierarchical processing, and in particular to a hierarchical processing system for energy storage data. Background Art

[0002] Modern energy storage data covers various types of equipment data (such as lithium batteries, lead-acid batteries, etc.) and is connected to complex energy networks. The generated data has multi-source heterogeneous characteristics, including equipment operating parameters, environmental monitoring data, user operation records, etc. Different data sources have significant differences in format, collection frequency and timestamp, and lack an effective fusion processing mechanism. It is difficult to form an organic whole, which seriously affects the accuracy and efficiency of data analysis. Traditional data classification methods are usually based on fixed rules or simple statistical features, which cannot adapt to the dynamic changes in energy storage data characteristics and business needs. In terms of weight determination, they cannot fully reflect the importance of data characteristics and their changes, which reduces the overall efficiency of data processing and increases the operation risk of energy storage systems.

[0003] At present, multi-source heterogeneous energy storage data cannot form an organic whole due to the lack of an effective association processing mechanism, and the traditional data classification system is difficult to dynamically adjust as data characteristics and business needs change, resulting in the classification results being out of touch with the importance and application scenarios of the actual data. This not only reduces the efficiency of data processing, but also may prevent key data from being processed and paid attention to in a timely manner, increasing the risk of energy storage system operation. Therefore, a hierarchical processing system for energy storage data is proposed herein. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention proposes the following technical solutions: A hierarchical processing system for energy storage data, comprising: Optimize feature extraction module: collect initial energy storage data, and use specified time intervals in the time dimension of the initial energy storage data Construct a unified time series for aligned sampling, introduce the initial energy storage data after aligned sampling into the autoencoder for noise reduction, perform feature extraction on the noise-reduced data, and obtain optimized energy storage feature data; Dynamic grading adjustment module: standardize the optimized energy storage feature data, calculate the information entropy of the standardized features to obtain the initial weights, introduce the Q reinforcement learning algorithm to optimize the weights online, design a three-order grading model based on the optimized initial weights, and perform data grading for different optimized energy storage feature data combinations and dynamic weights according to the three-order grading model; Hierarchical storage optimization module: Based on the hierarchical results of dynamic grading, a cost function is constructed. The storage cost change rate of each level of data at different time stages is calculated through the cost function, and the data heat index is calculated by combining the cost sensitivity coefficient and time weight, and a storage threshold is set. , when the data heat index exceeds the storage threshold When the data is frequently accessed, it is migrated to SSD. When the data heat index is lower than or equal to the storage threshold When the data is accessed infrequently, it is migrated to HDD.

[0005] The initial energy storage data includes energy storage equipment operation data, energy storage environment monitoring data and user operation record data; Assume that the total number of data points collected by all data sources is , collected The data point set of the data source is , forming multi-source heterogeneous initial energy storage data.

[0006] The process of aligning the data from each data source according to the unified time series is as follows: Align multi-source heterogeneous data from the time dimension, and Data points in , and its corresponding timestamp is , unify the data from all data sources into an optimal time interval Assume that The determination process is: Analyze and collect the original collection frequency of each data source. Data sources , and its original acquisition frequency is recorded as ,in For the Data sources The number of data points in ; Each data point has a timestamp. A time scale search range is set, starting from a time interval of 0.1 seconds, and gradually increasing to a time interval of 10 seconds with a certain step size of 0.1 per second. The time scale set is recorded as ,in is the minimum time scale, As the maximum time scale, obtain the change index of the data at each time scale, analyze the change trend of these change indexes with the time scale, and find the time scale that makes the data change index tend to be stable. The time interval corresponding to this time scale is .

[0007] The process of obtaining the optimized energy storage characteristic data is as follows: Based on time interval Constructing unified time series , is the starting sampling time point, As the last sampling time point, the data of each data source are aligned and sampled according to the unified time series to obtain the initial energy storage number after aligned sampling. ,in, is the number of all data points. After alignment, the number of overall data points remains unchanged. The initial energy storage data after aligned sampling is introduced into the autoencoder for noise reduction processing; The initial energy storage number after alignment sampling As the input of the autoencoder, the autoencoder maps the high-dimensional input data to a low-dimensional hidden layer representation through the encoder, and the decoder restores the hidden layer representation to the reconstructed data, and adjusts the encoder and decoder by minimizing the reconstruction error value of the reconstructed data. The formula is expressed as:

[0008] in, is the number of all input data, is the first data, is the first elements; The autoencoder is trained through the back-propagation algorithm, and the reconstruction error is passed back to each layer of the encoder and decoder. The gradient is calculated and the weight matrix and bias vector are updated so that the reconstruction error is gradually reduced. After training, the feature extraction of the denoised data is performed to obtain the optimized energy storage feature data. .

[0009] The optimized initial weight acquisition process is: Assume that the optimized energy storage characteristic data includes different feature dimensions, through a Optimized energy storage characteristic data matrix ,in To optimize the number of data points for energy storage characteristic data, In order to optimize the number of characteristic dimensions of energy storage characteristic data, each row corresponds to a data point, and each column corresponds to a characteristic dimension. The optimized energy storage characteristic data is standardized. feature dimensions, the standardized formula is:

[0010] in, Indicates The data points corresponding to the feature dimensions are It is The mean of the feature dimensions, It is The standard deviation of the feature dimension; Calculate the information entropy of each standardized feature. feature dimensions, information entropy The steps to obtain are as follows: For feature dimensions, calculate each data point in Probability in Dimensions , each in the The standardized feature values ​​of the data points in the dimension Divide by the sum of the standardized feature values ​​of all data points in this feature dimension , get each data point in Probability in Dimensions ; According to the calculated probability , further calculate the Information entropy of feature dimensions , the formula is:

[0011] Then based on information entropy Get the initial weight for each feature dimension , the complementary value of each feature dimension Divide by the sum of the complementary values ​​of all feature dimensions , the result is the initial weight of the feature dimension ; Then the Q reinforcement learning algorithm is introduced to optimize the weights online; Determine a state space based on the range of feature dimension values ; Determine an action space based on a set of adjustment operations on each feature weight ; Create and initialize a of Value Matrix ,in is the size of the state space, is the size of the action space, at each time step the current state is determined and The probability of moving from the action space Select an action from The probability of choosing The action with the largest value will adjust the weight of the selected action and give an immediate reward based on the processing result. , when the accuracy of the classification result improves, a positive reward is given, and when the accuracy decreases, a negative reward is given, and then the update value, repeat until the preset maximum number of learning steps is reached or The value matrix converges and finally obtains the optimized initial weights .

[0012] The implementation process of designing the three-order classification model based on the optimized initial weights is as follows: Construct a three-level classification model, the three levels of the model are emergency level, important level and ordinary level; Based on optimizing energy storage characteristic data Different feature dimensions and the initial weight vector optimized in the previous steps are recorded as ,in Indicates The optimized weights of feature dimensions ; The data is divided into three levels: emergency level , Importance and ordinary level , set two thresholds and , used as the basis for grading determination; Among them, emergency The judgment condition is: The feature dimension Data points ( ), and its optimized energy storage characteristic data is ,when When , the corresponding data point is judged as emergency level, where is the number of feature dimensions used for emergency level determination ; Importance The judgment conditions are: the emergency level conditions are not met and the , then the data point is judged as important; Normal level L3 judgment conditions: If the data point meets neither the emergency level conditions nor the important level conditions, it is judged as normal level; After the classification process by the three-order classification model, three groups of data were obtained, which are: , as well as .

[0013] The construction process of the cost function is: Suppose the storage medium set is ,in Represents SSD, Represents HDD, for storage media, storage Level, among which, =1, 2, 3, and correspond to , , , the unit storage cost of data is , the storage capacity is , the amount of data at each data level is , define a cost function to calculate the use of storage media Storage Storage cost of level data:

[0014] in, Represents the storage medium set as or The index of .

[0015] The process of obtaining the data heat index is as follows: The storage cost change rate of each level of data in different time stages is calculated through the cost function, and the storage cost change rate of each level of data in different time stages is calculated. The relative change ratio of the comprehensive storage cost of the first-level data is calculated by using The storage cost of the time period minus the storage cost of the The storage cost of a time period is calculated to obtain the change in storage cost, and then this change is divided by the storage cost of the previous time period to obtain the relative change rate. ; According to the importance of different data levels, set the cost sensitivity coefficient for each level ; Set a time weight according to different time periods , where each time stage There is a corresponding time weight , the sum of the weights of all time stages is 1; By storage cost change rate , cost sensitivity coefficient and time weight Calculate the data heat index: in, Indicates the bin to which the data point belongs.

[0016] The present invention has the following beneficial effects: In the present invention, firstly, by optimizing the feature extraction module, in terms of multi-source heterogeneous data processing, unified time series and aligned sampling solve the differences in format, acquisition frequency and timestamp between different data sources, and in the time dimension of the initial energy storage data, the time interval is used to obtain the energy storage data. A unified time series was constructed, and the data from each data source was aligned and sampled. The autoencoder was then used for noise reduction, which effectively solved the problem of lack of effective correlation processing mechanism for multi-source heterogeneous energy storage data. The originally scattered and formatted data were integrated into an organic whole, laying the foundation for subsequent data processing and analysis, and improving the availability and consistency of the data. Secondly, through the dynamic grading adjustment module, data standardization is used to eliminate the impact of dimensions, making data of different feature dimensions comparable, laying the foundation for the accurate calculation of information entropy and weights. Then, the initial weights are determined based on the information entropy to highlight important features. Features with low information entropy have high weights in subsequent analysis, which improves the accuracy and effectiveness of data processing. Then, the weights are optimized online through the Q reinforcement learning algorithm to adapt to the dynamic changes of data. Based on the optimized weights, a three-order grading model is constructed and the data is divided into emergency, important and ordinary levels, providing targeted processing and management strategies for data of different levels. Finally, through the hierarchical storage optimization module, according to the results of dynamic grading, a cost function is constructed to comprehensively consider the storage medium, data level, data volume and storage capacity, providing a quantitative basis for storage cost evaluation, assisting storage resource allocation decisions, and calculating the data heat index in combination with the cost sensitivity coefficient and time weight, which comprehensively reflects the changes in data heat and makes storage decisions more reasonable. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a system block diagram of a hierarchical processing system for energy storage data proposed by the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Embodiment 1 like Figure 1 As shown, the present invention proposes a hierarchical processing system for energy storage data, comprising: Optimize feature extraction module: collect initial energy storage data, and use specified time intervals in the time dimension of the initial energy storage data Construct a unified time series for aligned sampling, introduce the initial energy storage data after aligned sampling into the autoencoder for noise reduction, perform feature extraction on the noise-reduced data, and obtain optimized energy storage feature data; Optimize the implementation process of the feature extraction module: Collect initial energy storage data from different data sources, covering all aspects of the energy storage system operation process, including: Energy storage equipment operation data: collects equipment operation status data such as voltage, current, power, temperature of energy storage batteries, and power conversion parameters of inverters; Energy storage environment monitoring data: obtain environmental parameters such as ambient temperature, humidity, air pressure, and light intensity; User operation record data: records the user's operation instructions for the energy storage system, such as the time and power setting of charging and discharging operations; Assume that the total number of data points collected by all data sources is , collected The data point set of the data source is ,The data from these data sources differ in format, acquisition frequency, and timestamp, ,which constitutes multi-source heterogeneous initial energy storage data; Align multi-source heterogeneous data from the time dimension, and Data points in , and its corresponding timestamp is , unify the data from all data sources into an optimal time interval Assume that The determination process is: Analyze the original acquisition frequency of each data source. For example, the energy storage device operating sensor collects data multiple times per second, while the environmental monitoring device may collect data every few minutes. Collect the original acquisition frequency of each data source. Data sources , and its original acquisition frequency is recorded as ,in For the Data sources The number of data points in the time scale is 0.1, each data point has a timestamp, and a time scale search range is set, starting from a time interval of 0.1 seconds, and gradually increasing to a time interval of 10 seconds with a certain step size of 0.1 per second. The time scale set is recorded as ,in is the minimum time scale, As the maximum time scale, obtain the change index of the data at each time scale, analyze the change trend of these change indexes with the time scale, and find a time scale that makes the data change index tend to be stable. The time interval corresponding to this time scale is the appropriate value. ; Specifically, on the basis of analyzing the change trend, we look for a specific time scale so that the data change index tends to be stable. This is because when the time scale reaches this specific time scale, as the time scale increases further, the change amplitude of the data change index becomes very small, that is, it tends to be stable. This means that after this time scale, if the time interval continues to increase, the fluctuation characteristics of the data will not change significantly. Therefore, the time interval corresponding to this time scale is determined as the appropriate unified time interval. ; Determine the uniform time interval After that, construct the unified time series , is the starting sampling time point, As the last sampling time point, the data of each data source are aligned and sampled according to the unified time series to obtain the initial energy storage number after aligned sampling. ,in, is the number of all data points. After alignment, the number of overall data points remains unchanged. After aligning and sampling the data from each data source in a unified time series, the data from different data sources have consistency on the time axis, which is convenient for subsequent analysis and processing; The process of introducing the initial energy storage data after aligned sampling into the autoencoder for noise reduction is: The initial energy storage number after alignment sampling As the input of the autoencoder, the autoencoder maps the high-dimensional input data to a low-dimensional hidden layer representation through the encoder, and the decoder restores the hidden layer representation to the reconstructed data, and adjusts the encoder and decoder by minimizing the reconstruction error value of the reconstructed data. The formula is expressed as:

[0020] in, is the number of all input data (one input data corresponds to one initial energy storage data after alignment sampling, which is the same as the above same), is the first data, is the first elements; Then, through the back-propagation algorithm, the reconstruction error is passed back to each layer of the encoder and decoder, the gradient is calculated and the weight matrix and bias vector are updated, so that the reconstruction error gradually decreases. After training, the autoencoder can effectively remove the noise in the input data and extract purer and more representative data features, providing high-quality data for the subsequent energy storage data processing steps; After the autoencoder is trained by the back-propagation algorithm, feature extraction is performed on the noise-reduced data to obtain optimized energy storage feature data. .

[0021] Dynamic grading adjustment module: standardize the optimized energy storage feature data, calculate the information entropy of the standardized features to obtain the initial weights, introduce the Q reinforcement learning algorithm to optimize the weights online, design a three-order grading model based on the optimized initial weights, and perform data grading for different optimized energy storage feature data combinations and dynamic weights according to the three-order grading model; Implementation process of dynamic grading adjustment module: Assume that the optimized energy storage characteristic data includes different feature dimensions, through a Optimized energy storage characteristic data matrix ,in To optimize the number of data points for energy storage characteristic data, In order to optimize the number of characteristic dimensions of energy storage characteristic data, each row corresponds to a data point, and each column corresponds to a characteristic dimension. The optimized energy storage characteristic data is standardized. Feature Dimensions , the standardized formula is:

[0022] in, Indicates The data points corresponding to the feature dimensions are It is The mean of the feature dimensions, It is The standard deviation of the feature dimension; Calculate the information entropy of each standardized feature. feature dimensions, information entropy The steps to obtain are as follows: For feature dimensions, calculate each data point in Probability in Dimensions , each in the The standardized feature values ​​of the data points in the dimension Divide by the sum of the standardized feature values ​​of all data points in this feature dimension , get each data point in Probability in Dimensions ; According to the calculated probability , further calculate the Information entropy of feature dimensions , first for each data point in The probability of the feature dimension Take the natural logarithm and multiply it by itself, then sum the above results for each data point and multiply the sum by , thus obtaining Information entropy of feature dimensions , the formula is:

[0023] Then based on information entropy Get the initial weight for each feature dimension , the complementary value of each feature dimension Divide by the sum of the complementary values ​​of all feature dimensions , the result is the initial weight of the feature dimension ; Among them, the initial weight is assigned according to the information entropy of the feature dimension. The feature dimension with low information entropy (i.e., high effective information ratio) has The value of is relatively large, and the initial weight obtained after dividing by the sum This means that in the subsequent data processing and analysis, these feature dimensions will be given greater weight and can play their role more fully, while the feature dimensions with high information entropy (low effective information ratio) will be given more weight. The value of is relatively small, and the initial weight The weight distribution method is also relatively low, and the impact on subsequent operations is relatively small. This weight distribution method helps to highlight important features and improve the accuracy and effectiveness of data processing and analysis; The implementation process of introducing the Q reinforcement learning algorithm to optimize the weights online is as follows: Determine a state space based on the range of feature dimension values , an action space is determined based on the set of adjustment operations on each feature weight , create and initialize a of Value Matrix ,in is the size of the state space, is the size of the action space, at each time step the current state is determined and The probability of moving from the action space Select an action from The probability of choosing The action with the largest value will adjust the weight of the selected action and give an immediate reward based on the processing result. , when the accuracy of the classification result improves, a positive reward is given, and when the accuracy decreases, a negative reward is given, and then the update Repeat the above steps until the termination condition is met, and finally obtain the optimized initial weight ; Among them, under the framework of Q reinforcement learning algorithm, the action space Contains various adjustment operations for feature weights. When at a certain time step from the action space After selecting an action, you need to execute this action and modify the feature weight accordingly. Taking increasing the weight of a feature as an example, the adjustment range is pre-set (increasing the current weight of the feature by 10%). Assume that the current feature weight is , when the action selected is to increase the weight of the feature, the adjusted weight value will increase by 10% of the current weight; when the action selected is to increase the weight of the feature, the adjusted weight value will decrease by 10% of the current weight; Specifically, when selecting an action, The initial value can be set to 0.2, and it is dynamically adjusted according to the convergence during the learning process. The learning rate range is (0, 1), and it is recommended to set the initial value to 0.1. It gradually decays according to the learning effect. The discount factor range is (0, 1), and it is recommended to set the initial value to 0.9; Termination condition requirements: set to reach the preset maximum number of learning steps (1000 steps), The value matrix converges (between two consecutive iterations The absolute sum of the value changes is less than Or the data classification performance index reaches a satisfactory level (classification accuracy reaches more than 95%)); The implementation process of designing a three-order classification model based on the optimized initial weights is as follows: Construct a three-level classification model, the three levels of the model are emergency level, important level and ordinary level; Known optimized energy storage characteristic data include different feature dimensions, the initial weight vector optimized by the previous steps is recorded as ,in Indicates The optimized weights of feature dimensions ; The data is divided into three levels: emergency level , Importance and ordinary level , set two thresholds and , used as the basis for grading determination; Among them, emergency The judgment condition is: The feature dimension Data points ( ), and its optimized energy storage characteristic data is ,when When , the corresponding data point is judged as emergency level, where is the number of feature dimensions used for emergency level determination , determined according to actual business needs and data characteristics; Importance The judgment condition is: if the data point does not meet the emergency level condition, and meets , then the data point is judged as important; Normal level L3 judgment conditions: If the data point meets neither the emergency level conditions nor the important level conditions, it is judged as normal level; After the classification process by the three-order classification model, three groups of data were obtained, which are: , as well as ; Specifically, emergency level Refers to data closely related to the key operating status and potential risks of the energy storage system, emergency level The data feature dimensions are usually closely related to the critical state or emergency situation of the system. For example, in the energy storage system, the core parameters such as battery voltage and current will be included in the feature dimensions of emergency level judgment and given a higher weight. When these key feature dimensions exceed the threshold through the grading model, When , it indicates that the system is in an emergency state, and the corresponding data point is judged as emergency level; Importance Refers to data that reflects important operating parameters and related information of the energy storage system. For the importance level judgment, under the premise that the data point does not meet the emergency level conditions, consider whether the results obtained by the classification model of the remaining feature dimensions exceed the threshold ,Although this part of feature dimensions does not dominate in emergency judgment, it is crucial for reflecting the ,important operating parameters and related information of the system, such as ,ambient humidity, charging and discharging power and other features; Ordinary level L3 refers to general operation record data. If the data point meets neither the emergency level condition nor the important level condition, it is judged as ordinary level. That is, when neither of the above two judgment conditions is met, the data point is classified as ordinary level.

[0024] Hierarchical storage optimization module: Based on the hierarchical results of dynamic grading, a cost function is constructed. The storage cost change rate of each level of data at different time stages is calculated through the cost function, and the data heat index is calculated by combining the cost sensitivity coefficient and time weight, and a storage threshold is set. , when the data heat index exceeds the storage threshold When the data is frequently accessed, it is migrated to SSD. When the data heat index is lower than or equal to the storage threshold When the data is accessed infrequently, it is migrated to HDD; Implementation process of the hierarchical storage optimization module: According to the classification results of dynamic classification, a cost function is constructed, and the storage medium set is ,in Represents SSD, Represents HDD, for storage media, storage class( =1, 2, 3, and correspond to , , The unit storage cost of data is , the storage capacity is , the amount of data at each data level is , define a cost function to calculate the use of storage media ( =1, 2) store the Storage cost of level data: , Represents a collection of storage media or The index of The storage cost change rate of each level of data in different time stages is calculated through the cost function, and the storage cost change rate of each level of data in different time stages is calculated. The relative change ratio of the comprehensive storage cost of the first-level data is calculated by using The storage cost of the time period minus the storage cost of the The storage cost of a time period is calculated to obtain the change in storage cost, and then this change is divided by the storage cost of the previous time period to obtain the relative change rate. , At the same time, according to the importance of different levels of data, a cost sensitivity coefficient is set for each level , , for example, emergency level Sensitivity coefficient Set to , importance level Sensitivity coefficient Set to , General Sensitivity coefficient Set to ,The larger the sensitivity coefficient, the more sensitive the high-level data is to cost changes; Set a time weight according to different time periods , ( , Indicates the total number of time segments divided. ), each time stage There is a corresponding time weight , the sum of the weights of all time stages is 1; Finally, by storing the cost change rate , cost sensitivity coefficient and time weight Calculate the data heat index: in, Indicates the level where the data point belongs; Set a storage threshold ,when That is, the data heat index exceeds the upper threshold When the data is frequently accessed, it is migrated to SSD. That is, the data heat index is lower than or equal to the storage threshold When the data is accessed infrequently, it is migrated to HDD.

[0025] In this embodiment, when faced with the massive multi-source heterogeneous data generated by numerous energy storage devices and complex operating environments, through this solution, the feature extraction module is optimized to collect sensor data, environmental monitoring data and user operation records of each device, unify the time series and reduce noise, integrate the data of each energy storage node and provide high-quality data for power station operation and maintenance personnel after processing, the dynamic grading adjustment module grades the data, reflects the importance of the system status, and provides a basis for energy scheduling. According to the hierarchical storage optimization module, the operation and maintenance personnel can focus on emergency-level data, timely discover and deal with potential risks, and the hierarchical storage optimization module decides the data storage location based on the heat index and threshold, comprehensively considers the storage medium, data level, data volume and storage capacity, provides a quantitative basis for storage cost evaluation, and assists in storage resource allocation decisions.

[0026] In the application, several formulas involved are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent actual situation. Some coefficients or weights in the formulas are set by technicians in this field according to actual conditions, so they will not be elaborated here.

[0027] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0028] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A hierarchical processing system for energy storage data, characterized in that: include: Optimize feature extraction module: collect initial energy storage data, and use specified time intervals in the time dimension of the initial energy storage data Construct a unified time series for aligned sampling, introduce the initial energy storage data after aligned sampling into the autoencoder for noise reduction, perform feature extraction on the noise-reduced data, and obtain optimized energy storage feature data; Dynamic grading adjustment module: standardize the optimized energy storage feature data, calculate the information entropy of the standardized features to obtain the initial weights, introduce the Q reinforcement learning algorithm to optimize the weights online, design a three-order grading model based on the optimized initial weights, and perform data grading for different optimized energy storage feature data combinations and dynamic weights according to the three-order grading model; Hierarchical storage optimization module: Based on the hierarchical results of dynamic grading, a cost function is constructed. The storage cost change rate of each level of data at different time stages is calculated through the cost function, and the data heat index is calculated by combining the cost sensitivity coefficient and time weight, and a storage threshold is set. , when the data heat index exceeds the storage threshold When the data is frequently accessed, it is migrated to SSD. When the data heat index is lower than or equal to the storage threshold When the data is accessed infrequently, it is migrated to HDD.

2. A hierarchical processing system for energy storage data according to claim 1, characterized in that: The initial energy storage data includes energy storage equipment operation data, energy storage environment monitoring data and user operation record data; Assume that the total number of data points collected by all data sources is , collected The data point set of the data source is , forming multi-source heterogeneous initial energy storage data.

3. A hierarchical processing system for energy storage data according to claim 1, characterized in that: The process of aligning the data from each data source according to the unified time series is as follows: Align multi-source heterogeneous data from the time dimension, and Data points in , and its corresponding timestamp is , unify the data from all data sources into an optimal time interval Assume that The determination process is: Analyze and collect the original collection frequency of each data source. Data sources , and its original acquisition frequency is recorded as ,in For the Data sources The number of data points in ; Each data point has a timestamp. A time scale search range is set, starting from a time interval of 0.1 seconds, and gradually increasing to a time interval of 10 seconds with a certain step size of 0.1 per second. The time scale set is recorded as ,in is the minimum time scale, As the maximum time scale, obtain the change index of the data at each time scale, analyze the change trend of these change indexes with the time scale, and find the time scale that makes the data change index tend to be stable. The time interval corresponding to this time scale is .

4. The hierarchical processing system for energy storage data according to claim 1, characterized in that: The process of obtaining the optimized energy storage characteristic data is as follows: Based on time interval Constructing unified time series , is the starting sampling time point, As the last sampling time point, the data of each data source are aligned and sampled according to the unified time series to obtain the initial energy storage number after aligned sampling. ,in, is the number of all data points. After alignment, the number of overall data points remains unchanged. The initial energy storage data after alignment sampling is introduced into the autoencoder for noise reduction processing; The initial energy storage number after alignment sampling As the input of the autoencoder, the autoencoder maps the high-dimensional input data to a low-dimensional hidden layer representation through the encoder, and the decoder restores the hidden layer representation to the reconstructed data, and adjusts the encoder and decoder by minimizing the reconstruction error value of the reconstructed data. The formula is expressed as: in, is the number of all input data, is the first data, is the first elements; The autoencoder is trained through the back-propagation algorithm, and the reconstruction error is passed back to each layer of the encoder and decoder. The gradient is calculated and the weight matrix and bias vector are updated so that the reconstruction error is gradually reduced. After training, the feature extraction of the denoised data is performed to obtain the optimized energy storage feature data. .

5. The hierarchical processing system for energy storage data according to claim 1, characterized in that: The optimized initial weight acquisition process is: Assume that the optimized energy storage characteristic data includes different feature dimensions, through a Optimized energy storage characteristic data matrix ,in To optimize the number of data points for energy storage characteristic data, In order to optimize the number of characteristic dimensions of energy storage characteristic data, each row corresponds to a data point, and each column corresponds to a characteristic dimension. The optimized energy storage characteristic data is standardized. feature dimensions, the standardized formula is: in, Indicates The data points corresponding to the feature dimensions are It is The mean of the feature dimensions, It is The standard deviation of the feature dimension; Calculate the information entropy of each standardized feature. feature dimensions, information entropy The steps to obtain are as follows: For feature dimensions, calculate each data point in Probability in Dimensions , each in the The standardized feature values ​​of the data points in the dimension Divide by the sum of the standardized feature values ​​of all data points in this feature dimension , get each data point in Probability in Dimensions ; According to the calculated probability , further calculate the Information entropy of feature dimensions , the formula is: Then based on information entropy Get the initial weight for each feature dimension , the complementary value of each feature dimension Divide by the sum of the complementary values ​​of all feature dimensions , the result is the initial weight of the feature dimension ; Then the Q reinforcement learning algorithm is introduced to optimize the weights online; Determine a state space based on the range of feature dimension values ; Determine an action space based on a set of adjustment operations on each feature weight ; Create and initialize a of Value Matrix ,in is the size of the state space, is the size of the action space, at each time step the current state is determined and The probability of moving from the action space Select an action from The probability of choosing The action with the largest value will adjust the weight of the selected action and give an immediate reward based on the processing result. , when the accuracy of the classification result improves, a positive reward is given, and when the accuracy decreases, a negative reward is given, and then the update value, repeat until the preset maximum number of learning steps is reached or The value matrix converges and finally obtains the optimized initial weights .

6. The hierarchical processing system for energy storage data according to claim 1, characterized in that: The implementation process of designing the three-order classification model based on the optimized initial weights is as follows: Construct a three-level classification model, the three levels of the model are emergency level, important level and ordinary level; Based on optimizing energy storage characteristic data Different feature dimensions and the initial weight vector optimized in the previous steps are recorded as ,in Indicates The optimized weights of feature dimensions ; The data is divided into three levels: emergency level , Importance and ordinary level , set two thresholds and , used as the basis for grading determination; Among them, emergency The judgment condition is: The feature dimension Data points ( ), and its optimized energy storage characteristic data is ,when When , the corresponding data point is judged as emergency level, where is the number of feature dimensions used for emergency level determination ; Importance The judgment conditions are: the emergency level conditions are not met and the , then the data point is judged as important; Normal level L3 judgment conditions: If the data point meets neither the emergency level conditions nor the important level conditions, it is judged as normal level; After the classification process by the three-order classification model, three groups of data were obtained, which are: , as well as .

7. The hierarchical processing system for energy storage data according to claim 1, characterized in that: The construction process of the cost function is: Suppose the storage medium set is ,in Represents SSD, Represents HDD, for storage media, storage Level, among which, =1, 2, 3, and correspond to , , , the unit storage cost of data is , the storage capacity is , the amount of data at each data level is , define a cost function to calculate the use of storage media Storage Storage cost of level data: in, Represents the storage medium set as or The index of .

8. The hierarchical processing system for energy storage data according to claim 1, characterized in that: The process of obtaining the data heat index is as follows: The cost function is used to calculate the storage cost change rate of each level of data in different time stages, and the storage cost change rate of each level of data in two adjacent time stages is calculated. The relative change ratio of the comprehensive storage cost of the first-level data is calculated by using The storage cost of the time period minus the storage cost of the The storage cost of a time period is calculated to obtain the change in storage cost, and then this change is divided by the storage cost of the previous time period to obtain the relative change rate. ; According to the importance of different data levels, set the cost sensitivity coefficient for each level ; Set a time weight according to different time periods , where each time stage There is a corresponding time weight , the sum of the weights of all time stages is 1; By storage cost change rate , cost sensitivity coefficient and time weight Calculate the data heat index: in, Indicates the bin to which the data point belongs.

Images