A distributed energy storage device data generation method and system based on adaptive policy optimization and long sequence dynamic enhancement network

Abstract

The application discloses a kind of based on adaptive strategy optimization and long sequence dynamic enhancement network's distributed energy storage equipment data generation method and system, it is related to data generation technical field, including: obtaining distributed energy storage equipment data, constructs the simulation environment of distributed energy storage equipment operation;Adaptive strategy optimization agent is constructed and is trained in simulation environment, utilizes physical time sequence constraint evaluation function feedback optimization strategy, iteratively trains until the data sequence of agent output meets distributed energy storage data;The generated data sequence is input into long sequence dynamic enhancement network, and data is enhanced and corrected processing;The quality of data after processing is evaluated using evaluation index, according to downstream task demand, set multistage threshold value, generate available data, and according to whether to meet the condition parameter adjustment, finally generate distributed energy storage equipment data.The application generates distributed energy storage operation characteristics by adaptive strategy optimization, produces similar synthetic data with actual data set distribution.

Description

Technical Field

[0001] The present invention relates to the technical field of data generation, and more particularly to a method and system for generating data for a distributed energy storage device based on adaptive strategy optimization and long sequence dynamic enhancement network. Background Art

[0002] The research and application of distributed energy storage systems often requires large amounts of data for equipment performance evaluation, fault diagnosis, and system optimization. However, the actual data collected may be limited, or data under certain specific operating conditions may be difficult to obtain. For example, data collection for energy storage devices in extreme environments may be restricted, yet studying the performance of energy storage devices in these conditions is crucial. Therefore, an effective data generation method is needed to expand the dataset to meet the needs of various research and applications, making it particularly important to support the research and application of distributed energy storage devices.

[0003] Therefore, how to propose a distributed energy storage device data generation method and system based on adaptive strategy optimization and long sequence dynamic enhancement network, and to use data generation methods to expand the data set when data is limited to meet the needs of various research and applications is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0004] In light of this, the present invention provides a distributed energy storage device data generation method and system based on adaptive strategy optimization and a long-sequence dynamic reinforcement network. This adaptive strategy optimization mechanism directly links to the physical laws governing the operation of energy storage devices, such as Ohm's law and the monotonically decreasing RUL characteristic. This ensures that the generated data not only has a statistical distribution consistent with real data but also possesses practical physical meaning, enabling direct application in downstream tasks requiring high data reliability, such as equipment fault diagnosis and lifespan prediction. The device operating status feature set architecture of the long-sequence dynamic reinforcement network can efficiently process long-sequence data spanning thousands of time steps. By strengthening the spatiotemporal correlations between current, voltage, temperature, and RUL through matrix parameter learning, it can capture the cumulative effects of long-term charge and discharge cycles on battery temperature and remaining life. The generated data exhibits significantly better long-range dependency consistency than traditional recurrent neural networks (RNNs). The data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network can use limited real data to generate new data with similar distribution and characteristics, produce synthetic data with similar distribution to the actual data set, and perform well in downstream tasks. Through real-time feedback from evaluation indicators, the generation strategy and model parameters are adjusted, and continuous optimization of data quality can be achieved without human intervention, forming a full process automation from data generation, feature enhancement to quality control, significantly improving data production efficiency and stability. To provide support for the research and application of distributed energy storage equipment, in order to achieve the above purpose, the present invention adopts the following technical solutions:

[0005] A distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic reinforcement network, comprising:

[0006] Obtain distributed energy storage equipment data and build a simulation environment for the operation of distributed energy storage equipment;

[0007] Construct an adaptive strategy optimization agent, train the agent in a simulation environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterate the training until the agent outputs a data sequence that conforms to the distributed energy storage data.

[0008] The generated data sequence is input into the long sequence dynamic enhancement network to enhance and correct the data;

[0009] Evaluation indicators are used to assess the quality of processed data. Based on downstream task requirements, such as fault diagnosis requiring MSE_I < 0.05 and RUL prediction requiring MAE_RUL < 5, multi-level thresholds are set: such as MSE_I < 0.05 and MAE_RUL < 5, to generate usable data. Parameters are adjusted based on whether the conditions are met, and finally, qualified distributed energy storage equipment data is generated.

[0010] Optionally, acquiring the distributed energy storage device data includes: collecting in the form of a time series, and sampling voltage, current, temperature and RUL value at a preset frequency.

[0011] Optionally, the construction of a simulation environment for the operation of the distributed energy storage device includes: defining the device operation state range of the device, the physical environment conditions of the device, the device operation state feature set, and the control strategy instruction set.

[0012] Optionally, the physical timing constraint evaluation function feedback optimization strategy includes:

[0013] Design physical timing constraints and give positive rewards if the generated data conforms to physical laws:

[0014] Physical plausibility bonus:

[0015] ;

[0016] in, is the generated voltage, is the generated current, is the internal resistance of the energy storage device;

[0017] If the data is abnormal, a negative reward will be given:

[0018] RUL logic rewards:

[0019] ;

[0020] Total Rewards: , by calculating Feedback strategy optimization direction.

[0021] Optionally, the step of constructing an adaptive strategy to optimize an intelligent agent and training the intelligent agent in a simulated environment includes:

[0022] Based on the deep adaptive strategy optimization algorithm, the generation parameters are adjusted according to the current device operating status to generate a preliminary data sequence. The optimization strategy is fed back using the physical timing constraint evaluation function. It is iteratively trained until the agent masters the generation mode of distributed energy storage data. The loss function is calculated as follows:

[0023]

[0024] in, is the probability ratio of the new and old strategies, is the advantage function, is the crop factor, is the mean square error of the value function, for , Indicates that Clip to interval Inside, The weight hyperparameter of the value function loss balances the impact of policy loss and value function loss. By combining a composite physical timing constraint evaluation function with a loss function algorithm, the agent can dynamically adjust its generation strategy based on feedback from the device operating status in the simulated environment. This allows the agent to accurately capture the current and voltage fluctuation characteristics of distributed energy storage devices under different operating conditions (such as normal operation, overload, and extreme temperatures), addressing the lack of adaptability of traditional methods to complex environments.

[0025] Optionally, the iterative training until the agent outputs a data sequence that conforms to the distributed energy storage data includes:

[0026] S11: The agent observes the current operating status of the device ;

[0027] S12: Agent follows the policy network Current device operating status Processing, loading real data statistical features (such as voltage mean 3.7V, temperature standard deviation 5℃), initializing the environment state S0=(I0, V0, T0, RUL0), the agent strategy network Randomly initialize parameters to generate initial control strategy instructions , output the probability distribution of control strategy instructions. For continuous control strategy instruction sets, the control strategy instructions obey Gaussian distribution;

[0028] ;

[0029] in, is the mean value of the control strategy instruction, is the variance of the control strategy instruction, and the control strategy instruction is sampled from the distribution ;

[0030] S13: The agent executes the selected control strategy instructions in the environment , and have an impact on the environment;

[0031] S14: The environment returns the next device operating state according to the control strategy instructions executed by the agent And corresponding rewards ,award Used to measure the performance of the agent at the current time step;

[0032] S15: Data record the equipment operating status of the current time step , control strategy instructions ,award and the next device operating status Record as a tuple , construct the data sequence generated by adaptive strategy optimization, iterate S11-S14, and when the termination condition is met, stop the interaction process to obtain the complete data sequence.

[0033] Optionally, inputting the generated data sequence into a long sequence dynamic enhancement network to enhance and correct the data includes:

[0034] S21: normalize the generated data sequence, input the normalized current, voltage, temperature and RUL sequence into the long sequence dynamic enhancement network, extract sequence features, and perform enhancement or correction operations on the data;

[0035] The forward propagation of the long sequence dynamic enhancement network is expressed as the device operation status update and output calculation:

[0036] ;

[0037] in, is the hidden device operating state at time t, integrating historical information with current input, is the input at time t, A, B, C, D are the equipment operation status feature set matrices, and the sequence characteristics and rules are learned through training. is the noise term, and the model suppresses abnormal interference through learning. is the output, i.e. the enhanced or corrected data;

[0038] S22: Feature capture and enhancement matrices A and B capture long-range dependencies of the sequence and strengthen key features. If the temperature data has abnormal values ​​due to interference, the model will be updated. Based on the historical equipment operating status and normal mode, suppressing influence Return to a reasonable range;

[0039] S23: RUL sequence optimization, combining current, voltage, and temperature sequences , through matrices C and D and Linear transformation, output corrected RUL prediction .

[0040] Optionally, the evaluating the quality of the processed data using an evaluation indicator includes:

[0041] The quality of the generated data is evaluated using evaluation indicators. For current, voltage, and temperature, the error between the generated data and the actual measured value is evaluated. For current, voltage, and temperature, the mean square error is used:

[0042] ;

[0043] For RUL, the prediction accuracy is evaluated using the mean absolute error:

[0044] .

[0045] Optionally, adjusting parameters according to whether a condition is met includes:

[0046] Based on the evaluation indicators, the generated data is judged to see whether it meets the pre-set accuracy and reliability standards. According to the downstream task requirements, multi-level thresholds are set to generate usable data.

[0047] If satisfied, output the data; if not satisfied, return to the parameter adjustment step to re-optimize and set the error threshold: , ;in, 、 is the threshold.

[0048] Optionally, a distributed energy storage device data generation system based on adaptive strategy optimization and long sequence dynamic reinforcement network includes:

[0049] Acquisition module: used to obtain distributed energy storage equipment data;

[0050] Simulation environment construction module: used to build a simulation environment for the operation of distributed energy storage equipment;

[0051] Data sequence output module: used to build an adaptive strategy optimization module, train the intelligent agent in a simulated environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterative training is carried out until the intelligent agent outputs a data sequence that conforms to the distributed energy storage data;

[0052] Data processing module: used to input the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data;

[0053] Parameter adjustment module: used to evaluate the quality of processed data using evaluation indicators. According to the requirements of downstream tasks, such as fault diagnosis requiring MSE_I < 0.05 and RUL prediction requiring MAE_RUL < 5, multi-level thresholds are set: such as MSE_I < 0.05, MAE_RUL < 5, to generate usable data, and adjust parameters based on whether the conditions are met, and finally generate distributed energy storage equipment data that meets the conditions.

[0054] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a distributed energy storage device data generation method and system based on adaptive strategy optimization and long sequence dynamic enhancement network, which has the following beneficial effects:

[0055] The present invention proposes a distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network, including: obtaining distributed energy storage device data and constructing a simulation environment for the operation of distributed energy storage equipment; constructing an adaptive strategy optimization intelligent agent, training the intelligent agent in the simulation environment, and using a physical timing constraint evaluation function to feedback the optimization strategy, iteratively training until the intelligent agent outputs a data sequence that conforms to the distributed energy storage data; inputting the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data; using evaluation indicators to evaluate the quality of the processed data, and adjusting parameters according to whether the conditions are met. According to the requirements of downstream tasks, such as fault diagnosis requiring MSE_I<0.05 and RUL prediction requiring MAE_RUL<5, multi-level thresholds are set: such as MSE_I<0.05, MAE_RUL<5, to generate usable data, and finally generate distributed energy storage device data that meets the conditions.

[0056] The present invention achieves: (1) More accurate long-sequence modeling: The current, voltage, and temperature data of distributed energy storage devices have obvious time series characteristics, and long-term change trends need to be considered to predict RUL. The unique architecture of the long-sequence dynamic enhancement network enables it to efficiently process long-sequence data. Its device operation status feature set architecture can efficiently process long-sequence data of thousands of time steps. Through matrix parameter learning, it strengthens the spatiotemporal correlation between current, voltage, temperature, and RUL. Compared with traditional models, it can more accurately capture the long-range dependencies in the operation data of distributed energy storage devices, such as the impact of long-term charge and discharge cycles on battery temperature and life, thereby generating current, voltage, temperature, and RUL data that are more in line with actual conditions.

[0057] (2) Strong adaptability to dynamic environments: The operating environment of distributed energy storage equipment is complex and changeable, and is affected by many factors such as light intensity, temperature, and load demand. The adaptive strategy optimization mechanism is directly related to the physical laws of energy storage equipment operation (such as Ohm's law and RUL monotonically decreasing characteristics), ensuring that the generated data is not only statistically consistent with the real data, but also has actual physical meaning. It can be directly used for downstream tasks such as equipment fault diagnosis and life prediction that require high data reliability. The adaptive strategy optimization combined with the long sequence dynamic reinforcement network method enables the model to learn the changing laws of these environmental factors. By combining the composite physical time constraint evaluation function with the loss function algorithm, the intelligent agent can dynamically adjust the generation strategy according to the state feedback of the simulated environment, accurately capture the current and voltage fluctuation characteristics of distributed energy storage equipment under different working conditions (such as normal operation, overload, extreme temperature), and generate more adaptive equipment operation data accordingly, better simulate the current, voltage, temperature fluctuations under different working conditions and their impact on RUL, and solve the problem that traditional methods are not adaptable enough to complex environments.

[0058] (3) Data generation and optimization integration: Traditional data generation methods often lack effective optimization mechanisms, while adaptive strategy optimization combined with long-sequence dynamic reinforcement networks can integrate data generation and optimization processes. By setting a reward mechanism, adaptive strategy optimization can guide long-sequence dynamic reinforcement networks to generate data that better meets actual needs and quality standards. For example, for distributed energy storage device data, a physical timing constraint evaluation function can be designed based on the accuracy, consistency, and effectiveness of equipment failure prediction, so that the model can continuously optimize the generated data and improve the reliability and practicality of the data. By adjusting the generation strategy and model parameters through real-time feedback of evaluation indicators, continuous optimization of data quality can be achieved without human intervention, forming a full process automation from data generation, feature enhancement to quality control, and significantly improving data production efficiency and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0060] Figure 1 A schematic flow chart of a distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network provided by the present invention.

[0061] Figure 2 This is the structure diagram of the adaptive strategy optimization agent provided by the present invention.

[0062] Figure 3 This is a diagram of the long sequence dynamic enhancement network structure provided by the present invention.

[0063] Figure 4 This is a comparison chart of the real data and generated data provided by the present invention.

[0064] Figure 5 This is the PCA result diagram provided by the present invention.

[0065] Figure 6 This is the t-SNE result diagram provided by the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0067] The embodiment of the present invention discloses a distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network, such as Figure 1 Shown, including:

[0068] Obtain distributed energy storage equipment data and build a simulation environment for the operation of distributed energy storage equipment;

[0069] Construct an adaptive strategy optimization agent, train the agent in a simulation environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterate the training until the agent outputs a data sequence that conforms to the distributed energy storage data.

[0070] The generated data sequence is input into the long sequence dynamic enhancement network to enhance and correct the data;

[0071] Use evaluation indicators to assess the quality of processed data. According to the requirements of downstream tasks, such as fault diagnosis requiring MSE_I < 0.05 and RUL prediction requiring MAE_RUL < 5, set multi-level thresholds: such as MSE_I < 0.05, MAE_RUL < 5, generate usable data, and adjust parameters based on whether the conditions are met, and finally generate distributed energy storage equipment data that meets the conditions.

[0072] In a specific embodiment, a distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network is specifically implemented as follows:

[0073] Step 1: Use the public dataset (LG dataset) for the LG18650HG2 distributed energy storage system. The discharge cycles follow the drive cycles, namely, UDDS, LA92, US06, and other cycles, as well as combinations of these. Specifically, the experiment used discharge cycles at 0°C, 10°C, and 25°C. The data was collected as a time series, sampling voltage, current, temperature, and RUL values ​​at a 10 Hz frequency.

[0074] Step 2: Clarify data generation requirements, specify the data to be generated regarding current, voltage, temperature, and remaining useful life (RUL), determine the purpose of the data, such as training a distributed energy storage device monitoring model, and the requirements for data accuracy and time span.

[0075] Step 3: Define the environment. Build an environment that simulates the operation of distributed energy storage devices, define the device's operating state range (e.g., normal operation, overload, etc.), and determine the device's physical environmental conditions (e.g., temperature range, humidity, etc.). The device operating state feature set can include various device operating state parameters, and the control strategy instruction set can be the device's operating instructions (e.g., adjusting voltage and current). Define the device operating state feature set S to include device operating state variables such as current I, voltage V, and temperature T, such as S = {(I, V, T, ...)}. The control strategy instruction set A contains device operation control strategy instructions, such as current, voltage, and temperature adjustment instructions.

[0076] Step 4: Design a physical timing constraint evaluation function to measure the agent's performance in operating the device in the simulated environment. For example, positive rewards are given if the device's operating parameters are within the normal range and close to the optimal value; negative rewards are given if the device exhibits an anomaly (such as overcurrent or overheating).

[0077] (1) Design reward rules. For example, positive rewards are given when the generated data conforms to physical laws (rationality of voltage and current correlation, temperature change trend). The following is the design formula of the physical timing constraint evaluation function:

[0078] Physical plausibility bonus:

[0079] ;

[0080] in, is the generated voltage, is the generated current, is the internal resistance of the energy storage device.

[0081] (2) If the data is abnormal (exceeding the device threshold, RUL logic contradiction), a negative reward is given to guide the adaptive strategy to optimize the agent to generate valid data.

[0082] RUL logic reward (ensuring monotonically decreasing remaining lifespan):

[0083] ;

[0084] (3) Total Rewards: ;

[0085] By calculation Feedback strategy optimization direction to ensure that the generated data conforms to physical laws and RUL logic.

[0086] Step 5: Figure 2 As shown in the figure, the adaptive strategy optimizes the training of the intelligent agent, allowing the intelligent agent to explore in a defined environment and generate training data through the device operating state-control strategy instruction sequence. Based on the deep adaptive strategy optimization algorithm, the generation parameters are adjusted according to the current device operating state to generate a preliminary data sequence. The optimization strategy is fed back using the physical timing constraint evaluation function. Iterative training is carried out until the intelligent agent masters the generation mode of distributed energy storage data. The loss function is calculated as follows:

[0087]

[0088] in, is the probability ratio of the new and old strategies, is the advantage function, is the crop factor, is the mean square error of the value function, for , Indicates that Clip to interval Inside, It is the weight hyperparameter of the value function loss, balancing the influence of policy loss and value function loss.

[0089] Step 6: At each time step, the agent interacts with the environment by following these steps:

[0090] Equipment operation status observation: The agent observes the current environment equipment operation status , which is the basis for the agent to make decisions.

[0091] Control strategy instruction selection: the agent selects the strategy according to the strategy network Current device operating status Processing, loading real data statistical features (such as voltage mean 3.7V, temperature standard deviation 5℃), initializing the environmental state =( , , , RU ). Agent Policy Network Randomly initialize parameters to generate initial control strategy instructions . Output the probability distribution of control strategy instructions.

[0092] For continuous control strategy instruction sets, it is often assumed that the control strategy follows a Gaussian distribution:

[0093] ;

[0094] in, is the mean value of the control strategy instruction, is the variance of the control strategy instruction. Then sample from this distribution to get the specific control strategy instruction .

[0095] Control strategy instruction execution: The agent executes the selected control strategy instructions in the environment , the control strategy instructions will have an impact on the environment.

[0096] Environmental feedback: The environment returns the next device operating state based on the control strategy instructions executed by the agent And corresponding rewards .award It is used to measure the performance of the agent at that time step, and its design is usually related to the task goal.

[0097] The state of the agent is (current, voltage, temperature, RUL), the control strategy instruction is (charge and discharge power, equipment adjustment instruction), and the data record will record the equipment operation status of the current time step , control strategy instructions ,award and the next device operating status Record as a tuple These tuples constitute the data sequence generated by the adaptive strategy optimization.

[0098] The termination condition can be reaching a preset number of time steps, completing a specific task, or the environment reaching a certain terminal state. When the termination condition is met, the interaction process stops and the complete data sequence is obtained.

[0099] Step 7: Long-sequence dynamic reinforcement network training. The value ranges of different features in the sequence data generated by adaptive strategy optimization may vary significantly. For example, the numerical ranges of current, voltage, temperature, and RUL vary. To prevent this scale difference from adversely affecting the long-sequence dynamic reinforcement network training, the data must first be normalized.

[0100] The generated current, voltage, temperature and RUL sequences are input into the long sequence dynamic enhancement network, and its efficient sequence processing capability is used to further extract sequence features and perform operations such as enhancement or correction on the data.

[0101] The long sequence dynamic enhancement network is based on the equipment operation status feature set model (SSM) architecture, which enhances or modifies the current, voltage, temperature and RUL sequences, such as Figure 3 As shown:

[0102] The basic formula of the equipment operation status feature set model is as follows:

[0103] The forward propagation of the long sequence dynamic enhancement network can be expressed as the device operation status update and output calculation:

[0104] ;

[0105] in, It is the hidden state at time t, integrating historical information with the current input. is the input at time t (such as a sequence value of current, voltage, temperature, etc.). A, B, C, and D are the device operating status feature set matrices, and the sequence characteristics and patterns are learned through training. is a noise term, and the model can suppress abnormal interference through learning. is the output, i.e. the enhanced or corrected data (e.g. the corrected RUL prediction value).

[0106] (1) Feature capture and enhancement: Matrices A and B capture long-range dependencies of the sequence (such as the impact of long-term temperature changes on RUL) and strengthen key features. For example, if a weak fluctuation in a certain period of time in the current sequence reflects the health status of the battery, the model amplifies this feature by adjusting matrix B. More prominently, finally Enhance the expression of this feature.

[0107] (2) Abnormal correction: if the temperature data has abnormal values ​​due to interference, the model will be updated Based on the historical equipment operating status and normal mode (encoded by matrices A, B), suppressing influence Return to a reasonable range.

[0108] (3) RUL sequence optimization, combining current, voltage, and temperature sequences , through matrices C and D and Linear transformation, output corrected RUL prediction For example, if historical data shows a strong correlation between voltage drop and RUL shortening, the model can strengthen this correlation by adjusting the matrix C and correct the part of the RUL prediction that does not conform to this pattern.

[0109] By training the device operating status feature set matrices A, B, C, and D, the intrinsic patterns of current, voltage, temperature, and RUL sequences are adaptively learned to achieve data enhancement (highlighting key features) and correction (suppressing anomalies and optimizing logical consistency).

[0110] Step 8: Use specific evaluation indicators to evaluate the quality of the generated data. For current, voltage, and temperature, the error between the generated data and the actual measured value can be evaluated. For current, voltage, and temperature, the mean square error (MSE) is used:

[0111] ;

[0112] For RUL, the prediction accuracy is evaluated using the mean absolute error:

[0113] .

[0114] Step 9: Determine whether the conditions are met. Based on the evaluation indicators, determine whether the generated data meets the pre-set standards such as accuracy and reliability. According to the requirements of downstream tasks, such as fault diagnosis requires MSE_I < 0.05 and RUL prediction requires MAE_RUL < 5, set multiple thresholds: such as MSE_I < 0.05, MAE_RUL < 5, to generate the final usable data. If it is met, output the data; if it is not met, return to the parameter adjustment step for re-optimization. Set the error threshold:

[0115] , ;in, 、 is the threshold.

[0116] Step 10: Adjust parameters. When the generated data does not meet the conditions, first locate the parameters and determine the core parameters of the agent that need to be adjusted. According to the specific problems that the data does not meet, lock the corresponding parameter categories:

[0117] 1. Adaptive strategy optimization strategy parameter adjustment (directly affects the control strategy instruction selection)

[0118] (1) Adjust the exploration rate If the data fluctuates too much (such as voltage surges and drops), it may be that the agent is over-exploring, resulting in unstable control strategy instructions. (For example, from 0.3 to 0.2), reduce the proportion of random control strategy instructions, and enhance the utilization of existing optimal strategies.

[0119] (2) Adjust the weights of the policy network. If the RUL prediction is continuously too low or too high, it may be that the policy network is mapping the device aging characteristics incorrectly. By backpropagating the gradient, the error signal is used to update , the formula is:

[0120] ;

[0121] in, is the policy loss function with respect to the parameters The gradient, is the learning rate, is the policy gradient loss function (such as the CLIP loss of the loss function algorithm).

[0122] Adjust the agent's parameters (such as the weights of the neural network) and training hyperparameters (such as the learning rate) based on the agent's performance during training, such as changes in reward values ​​and the accuracy of predicted RUL:

[0123] ;

[0124] in, is the current learning rate, is the decay coefficient, and t is the number of training steps.

[0125] 2. Adjustment of parameters of environmental and physical timing constraint evaluation functions (affecting device operation status-control strategy instruction mapping)

[0126] (1) Adjust the weight coefficient of the physical timing constraint evaluation function (such as the priority of current / voltage / temperature). If the temperature data exceeds the safety range for a long time, it may be that the temperature weight in the physical timing constraint evaluation function is too low. Too low. Increase (e.g. from 0.2 to 0.5), making the agent pay more attention to the temperature constraint, the formula is:

[0127] ;

[0128] in, is the voltage weight coefficient, is the RUL weight coefficient, is the reward component corresponding to RUL.

[0129] (2) Adjust the normalization parameters of the device operating state feature set (such as the current / voltage scaling factor). If the agent is not sensitive to changes in the device operating state (such as small voltage fluctuations do not trigger control strategy instructions), it may be over-normalization. Recalculate the normalized mean and standard deviation , or dynamic normalization, such as batch normalization.

[0130] 3. Long sequence dynamic enhancement network hyperparameter adjustment (influence sequence feature extraction)

[0131] (1) Adjust the number of attention heads or layers. If the long-period RUL prediction error is large, it may be that the long sequence dynamic enhancement network is insufficient to model long-term dependencies. Increase the number of layers (e.g., from 2 layers to 3 layers) or expand the hidden layer dimension (e.g., from 512 to 768) to enhance the ability of long sequence feature capture.

[0132] (2) Adjust the time step (time resolution of data generation), if the time granularity of generated data does not match the actual sampling frequency of the device (e.g., minute level vs. second level), resulting in error accumulation. Adjustment: match the actual sampling frequency, such as changing from 10 minutes to 1 minute, and retrain the agent to adapt to the new time step.

[0133] 4. Gradient-based update strategy adjustment

[0134] Gradient-driven parameter fine-tuning (online real-time adjustment) is suitable for scenarios where data errors are small and only the weight parameters of the strategy network or long sequence dynamic enhancement network need to be fine-tuned. Specific steps:

[0135] Input the unqualified data into the agent, calculate the deviation between the control strategy instruction output of the current strategy and the expected control strategy instruction (e.g., the expected voltage is , the actual generated voltage is ). Update the trainable parameters (such as weights W and biases b) of the strategy network or long sequence dynamic enhancement network through error backpropagation, the formula is:

[0136] ;

[0137] ;

[0138] where, is the learning rate, is the joint loss of data error and physical timing constraint evaluation function:

[0139] ;

[0140] After adjustment, repeat steps 5-10 again to generate data, and continue to repeat the above steps until the output data meets the requirements.

[0141] Step 11: Output data, when the generated data meets the conditions.

[0142] In a specific embodiment, a distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic reinforcement network is described. The specific comparative experiments are as follows:

[0143] (1) Comparative analysis of the single long sequence dynamic enhancement network, adaptive strategy optimization and this embodiment, such as Figure 4 The following figure shows an example of real data and a comparison of data generated by the three methods. The third line with relatively small fluctuations represents voltage, the second line with relatively large fluctuations represents current, the top line with relatively flatter fluctuations represents temperature, and the bottom line with relatively flatter fluctuations represents RUL. All features were rescaled using min-max normalization so that they all ranged between 0 and 1. This example shows that it better matches real data.

[0144] (2) PCA and t-SNE were used to analyze the data distribution comparison of real data and synthetic data. PCA and t-SNE were used to analyze the data distribution of real data and synthetic data. PCA and t-SNE were used to simplify the distributed energy storage operation data series with four features into two-dimensional data so that their distribution can be visualized in two dimensions. The output of PCA and t-SNE was generated by the machine learning library sklearn. The PCA results are shown in Figure 2. Figure 5 As shown, the t-SNE results are as follows Figure 6 As shown in the figure. Each point in the figure represents a sequence of average values ​​of four characteristics (voltage, current, temperature, and RUL). The scattered points on the periphery represent real data, while the more densely packed points in the middle are generated data. If the data points overlap, the distribution is considered close. Based on their overlapping area, it can be obtained that this embodiment can better approximate the real data distribution.

[0145] In a specific embodiment, a distributed energy storage device data generation system based on adaptive strategy optimization and long sequence dynamic enhancement network includes:

[0146] Acquisition module: used to obtain distributed energy storage equipment data;

[0147] Simulation environment construction module: used to build a simulation environment for the operation of distributed energy storage equipment;

[0148] Data sequence output module: used to build an adaptive strategy optimization agent, train the agent in a simulation environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterative training is carried out until the agent outputs a data sequence that conforms to the distributed energy storage data.

[0149] Data processing module: used to input the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data;

[0150] Parameter adjustment module: used to evaluate the quality of processed data using evaluation indicators. According to the requirements of downstream tasks, such as fault diagnosis requiring MSE_I < 0.05 and RUL prediction requiring MAE_RUL < 5, multi-level thresholds are set: such as MSE_I < 0.05, MAE_RUL < 5, to generate usable data, and adjust parameters based on whether the conditions are met, and finally generate distributed energy storage equipment data that meets the conditions.

[0151] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0152] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network, characterized in that: include: Obtain distributed energy storage equipment data and build a simulation environment for the operation of distributed energy storage equipment; Construct an adaptive strategy optimization agent, train the agent in a simulation environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterate the training until the agent outputs a data sequence that conforms to the distributed energy storage data. The construction of the adaptive strategy optimization agent and the training of the agent in the simulation environment include: Based on the deep adaptive strategy optimization algorithm, the generation parameters are adjusted according to the current device operating status to generate a preliminary data sequence. The optimization strategy is fed back using the physical timing constraint evaluation function. It is iteratively trained until the agent masters the generation mode of distributed energy storage data. The loss function is calculated as follows: J(θ)=E k [min(p k (i)A k ,clip(p k (θ), 1-δ, 1+δ)A k )-αVFloss]; Among them, p k (θ) is the probability ratio of the new and old strategies, A k is the advantage function, δ is the clipping coefficient, VF loss is the mean square error of the value function, and clip is clip(p k (θ), 1-δ, 1+δ), which means p k (θ) is clipped to the interval (1-δ, 1+δ), α is the weight hyperparameter of the value function loss, balancing the influence of policy loss and value function loss; The iterative training until the agent outputs a data sequence that conforms to the distributed energy storage data includes: S11: The agent observes the current device operating status S t ; S12: Agent follows the policy network π θ (a|s) for the current device operating status S t Process, load real data statistical features, initialize the environment state S0, and the agent strategy network π θ (a|s) Randomly initialize parameters to generate initial control strategy instructions a t , output the probability distribution of control strategy instructions. For continuous control strategy instruction sets, the control strategy instructions obey Gaussian distribution; p θ (a|s)=N(μ θ (s),s θ 2 (s)); Among them, μ θ (s) is the mean value of the control strategy instruction, σ θ 2 (s) is the variance of the control strategy instruction, and the control strategy instruction a is sampled from the distribution t ; S13: The agent executes the selected control strategy instruction a in the environment t , and have an impact on the environment; S14: The environment returns the next device operation state s according to the control strategy instructions executed by the agent t+1 And the corresponding reward r t , reward r t Used to measure the performance of the agent at the current time step; S15: Data record the equipment operation status s at the current time step t , control strategy instruction a t , reward r t and the next device operating state s t+1 Record as a tuple (s t , a t , r t , s t+1 ), construct the data sequence generated by adaptive strategy optimization, iterate S11-S14, and when the termination condition is met, stop the interaction process to obtain a complete data sequence; The generated data sequence is input into the long sequence dynamic enhancement network to enhance and correct the data; Inputting the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data includes: S21: normalize the generated data sequence, input the normalized current, voltage, temperature and RUL sequence into the long sequence dynamic enhancement network, extract sequence features, and perform enhancement or correction operations on the data; The forward propagation of the long sequence dynamic enhancement network is expressed as the device operation status update and output calculation: Among them, S t is the hidden device operating state at time t, integrating historical information with current input, X t is the input at time t, A, B, C, D are the equipment operation status feature set matrices, and the sequence features and rules are learned through training, ∈ t is the noise term, and the model suppresses abnormal interference through learning. t is the output, i.e. the enhanced or corrected data; S22: Feature capture and enhancement matrices A and B capture long-range dependencies of the sequence and strengthen key features. If the temperature data has abnormal values ​​due to interference, the model will update S t When, based on the historical equipment operating status S t-1 and normal mode, suppressing ∈ t Influence, make Y t Return to a reasonable range; S23: RUL sequence optimization, combining current, voltage, and temperature sequences x t , through matrices C and D on S t and X t Linear transformation, output corrected RUL prediction Y t ; Use evaluation indicators to assess the quality of processed data, set multi-level thresholds according to downstream task requirements, generate usable data, and adjust parameters based on whether the conditions are met, ultimately generating qualified distributed energy storage device data.

2. A distributed energy storage device data generation method based on adaptive strategy optimization and long sequence dynamic enhancement network according to claim 1, characterized in that: The obtaining of distributed energy storage device data includes: collecting in the form of time series, sampling voltage, current, temperature and RUL value at a preset frequency.

3. The method for generating distributed energy storage device data based on adaptive strategy optimization and long sequence dynamic enhancement network according to claim 1, characterized in that: The construction of the simulation environment for the operation of the distributed energy storage device includes: defining the working state range of the device, the physical environment conditions of the device, the device operation state feature set and the control strategy instruction set.

4. The method for generating distributed energy storage device data based on adaptive strategy optimization and long sequence dynamic enhancement network according to claim 1, characterized in that: The physical timing constraint evaluation function feedback optimization strategy includes: Design physical timing constraints and give positive rewards if the generated data conforms to physical laws: Physical plausibility bonus: R phys =-|V gen -I gen ×R|; Among them, V gen is the generated voltage, I gen is the generated current, R is the internal resistance of the energy storage device; If the data is abnormal, a negative reward will be given: RUL logic rewards: Total Reward: R total =R phys +R rul , by calculating R total Feedback strategy optimization direction.

5. The method for generating distributed energy storage device data based on adaptive strategy optimization and long sequence dynamic enhancement network according to claim 1, characterized in that: The evaluation indicators used to evaluate the quality of processed data include: Use evaluation indicators to evaluate the quality of generated data. For current, voltage, and temperature, use mean square error to evaluate the error between them and the actual measured values: For RUL, the prediction accuracy is evaluated using the mean absolute error:

6. The method for generating distributed energy storage device data based on adaptive strategy optimization and long sequence dynamic enhancement network according to claim 1, characterized in that: The parameter adjustment according to whether the conditions are met includes: Based on the evaluation indicators, the generated data is judged to see whether it meets the pre-set accuracy and reliability standards. According to the downstream task requirements, multi-level thresholds are set to generate usable data. If satisfied, output data; if not satisfied, return to the parameter adjustment step to re-optimize and set the error threshold: MSE I <ξ1,MSE RUL <ξ2; where ξ1 and ξ2 are thresholds.

7. A distributed energy storage device data generation system based on adaptive strategy optimization and long sequence dynamic enhancement network, characterized in that: include: Acquisition module: used to obtain distributed energy storage equipment data; Simulation environment construction module: used to build a simulation environment for the operation of distributed energy storage equipment; Data sequence output module: used to build an adaptive strategy optimization module, train the intelligent agent in a simulated environment, and use the physical timing constraint evaluation function to feedback the optimization strategy. Iterative training is carried out until the intelligent agent outputs a data sequence that conforms to the distributed energy storage data; Building an adaptive strategy optimization agent and training the agent in a simulated environment includes: Based on the deep adaptive strategy optimization algorithm, the generation parameters are adjusted according to the current device operating status to generate a preliminary data sequence. The optimization strategy is fed back using the physical timing constraint evaluation function. It is iteratively trained until the agent masters the generation mode of distributed energy storage data. The loss function is calculated as follows: J(θ)=E k [min(p k (i)A k ,clip(p k (θ), 1-δ, 1+δ)A k )-αVFloss]; Among them, p k (θ) is the probability ratio of the new and old strategies, A k is the advantage function, δ is the clipping coefficient, VF loss is the mean square error of the value function, and clip is clip(p k (θ), 1-δ, 1+δ), which means clipping pk(θ) to the interval (1-δ, 1+δ), α is the weight hyperparameter of the value function loss, balancing the influence of policy loss and value function loss; The iterative training until the agent outputs a data sequence that conforms to the distributed energy storage data includes: S11: The agent observes the current device operating status S t ; S12: Agent follows the policy network π θ (a|s) for the current device operating status S t Process, load real data statistical features, initialize the environment state S0, and the agent strategy network π θ (a|s) Randomly initialize parameters to generate initial control strategy instructions a t , output the probability distribution of control strategy instructions. For continuous control strategy instruction sets, the control strategy instructions obey Gaussian distribution; p θ (a|s)=N(μ θ (s),s θ 2 (s)); Among them, μ θ (s) is the mean value of the control strategy instruction, σ θ 2 (s) is the variance of the control strategy instruction, and the control strategy instruction a is sampled from the distribution t ; S13: The agent executes the selected control strategy instruction a in the environment t , and have an impact on the environment; S14: The environment returns the next device operation state s according to the control strategy instructions executed by the agent t+1 And the corresponding reward r t , reward r t Used to measure the performance of the agent at the current time step; S15: Data record the equipment operation status s at the current time step t , control strategy instruction a t , reward r t and the next device operating state s t+1 Record as a tuple (s t , a t , r t , s t+1 ), construct the data sequence generated by adaptive strategy optimization, iterate S11-S14, and when the termination condition is met, stop the interaction process to obtain a complete data sequence; Data processing module: used to input the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data; Inputting the generated data sequence into the long sequence dynamic enhancement network to enhance and correct the data includes: S21: normalize the generated data sequence, input the normalized current, voltage, temperature and RUL sequence into the long sequence dynamic enhancement network, extract sequence features, and perform enhancement or correction operations on the data; The forward propagation of the long sequence dynamic enhancement network is expressed as the device operation status update and output calculation: Among them, S t is the hidden device operating state at time t, integrating historical information with current input, X t is the input at time t, A, B, C, D are the equipment operation status feature set matrices, and the sequence features and rules are learned through training, ∈ t is the noise term, and the model suppresses abnormal interference through learning. t is the output, i.e. the enhanced or corrected data; S22: Feature capture and enhancement matrices A and B capture long-range dependencies of the sequence and strengthen key features. If the temperature data has abnormal values ​​due to interference, the model will update S t When, based on the historical equipment operating status S t-1 and normal mode, suppressing ∈ t Influence, make Y t Return to a reasonable range; S23: RUL sequence optimization, combining current, voltage, and temperature sequences x t , through matrices C and D on S t and X t Linear transformation, output corrected RUL prediction Y t ; Parameter adjustment module: used to evaluate the quality of processed data using evaluation indicators, set multi-level thresholds according to downstream task requirements, generate usable data, and adjust parameters based on whether the conditions are met, ultimately generating distributed energy storage device data that meets the conditions.

Images