Time domain decision generation method for active retirement of power battery based on diffusion reinforcement learning

Abstract

This invention proposes a temporal decision generation method for proactive decommissioning of power batteries based on diffusion reinforcement learning. The method includes: constructing a state vector for the predicted state; calculating the benefit-loss ratio sequence for each time point within a future period and identifying the benefit window with the greatest future usability; calculating a future risk score to predict irreversible degradation windows where the battery may experience irreversible performance decline; generating multiple future strategy trajectories containing future states and decommissioning actions using a diffusion model, and optimizing the trajectories by combining the benefit-loss ratio sequence and the irreversible degradation window using reinforcement learning to select the optimal decommissioning trajectory; and generating executable behavioral suggestions, including immediate decommissioning, delayed observation, or continued operation, based on the recommended decommissioning time and corresponding state determined by the optimal decommissioning trajectory and a preset behavioral evaluation logic. This invention is applicable to the proactive decommissioning management of vehicle-mounted power batteries, energy storage systems, and multi-module battery packs.

Description

Technical Field

[0001] This invention belongs to the technical field of time-domain decision generation for active decommissioning of power batteries, and particularly relates to a method for generating time-domain decisions for active decommissioning of power batteries based on diffusion reinforcement learning. Background Technology

[0002] With the increasing prevalence of new energy vehicles, distributed energy storage, and grid-side energy storage systems, the scale of in-service power batteries is growing rapidly, making proactive retirement management increasingly important. However, power batteries are affected by multiple factors during service, including load fluctuations, temperature environment, and charging / discharging strategies, exhibiting complex time-series degradation characteristics. Their performance degradation is not linear but may accelerate or even reach critical irreversible degradation in certain segments. Most existing retirement strategies rely on static thresholds for retirement determination, such as single-point indicators like health percentage, internal resistance limit, or cycle count, which fail to reflect the changing trend of the battery's future usability. Furthermore, these methods cannot identify potential acceleration windows for battery degradation. In some cases, even if current health indicators are acceptable, a rapid increase in temperature or abnormal load in the future may quickly lead to irreversible damage, increasing safety risks. On the other hand, with the increasing economic demands of applications such as vehicle operation and peak-valley electricity pricing, the benefits of continuing to use batteries exhibit significant time-varying characteristics. Fixed-threshold retirement strategies struggle to balance maximizing benefits with controlling safety risks. In recent years, although some studies have attempted to improve the scientific nature of retirement strategies by using lifespan prediction or machine learning, these methods still rely on the idea of ​​formulating fixed criteria based on prediction results. They lack the ability to generate future strategy sequences and have failed to model changes in benefits, degradation trends, dynamic risks, and strategic behaviors in a unified manner.

[0003] Therefore, in complex operating environments, there is a need for a novel proactive decommissioning method that can combine multi-source state information, dynamic economic benefits, future degradation risks, and generative behavioral strategies, so as to achieve optimal usage planning for power batteries while avoiding entering the irreversible damage range. Summary of the Invention

[0004] The purpose of this invention is to propose a time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning, so as to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention provides a time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning, comprising the following steps: S1. Construct a state vector for the predicted state; wherein, the state vector includes the health status, internal resistance, voltage, current, temperature trend, load power, unit benefit prediction, and performance degradation rate of the power battery; S2. Based on the unit benefit prediction and performance degradation rate, calculate the benefit-loss ratio sequence at each time point in the future, and identify the benefit window with the greatest future use value. S3. Based on the state vector and benefit-loss ratio sequence, and taking into account impedance growth, temperature change, voltage change, load and benefit factor, calculate the future risk score to predict the irreversible degradation window in which the battery may experience irreversible performance degradation. S4. Using the state vector as a condition, generate multiple future policy trajectories containing future states and retirement actions using a diffusion model, and use reinforcement learning combined with the benefit-loss ratio sequence and the irreversible degradation window to score and optimize the trajectories, and select the optimal retirement trajectory. S5. Based on the recommended retirement time and its corresponding status determined by the optimal retirement trajectory, and combined with the preset behavior evaluation logic, generate executable behavior suggestions including immediate retirement, delayed observation, or continued operation.

[0006] Furthermore, the performance degradation rate is calculated by the time-series evolution trend of the internal resistance.

[0007] Furthermore, the benefit-loss ratio of the benefit-loss ratio sequence at each time point in the future period is generated by calculating the predicted unit benefit and performance degradation rate at each time point.

[0008] Furthermore, when calculating the benefit-loss ratio at each moment in the future period, a stability adjustment term for the second-order change of the unit benefit prediction value is introduced to correct the error, which is used to suppress the misleading effect of short-term high fluctuation range on decision-making, and finally obtain the benefit-loss ratio sequence at each moment in the future period.

[0009] Furthermore, the benefit factor is calculated and generated based on the benefit-loss ratio sequence at each time point within the future period.

[0010] Furthermore, S3 also includes: A standardized risk score is calculated for each time point within the future interval, and this future risk score is mapped to... Risk probability value within the range; The set of time periods in which the risk probability value is greater than a preset threshold is used as the irreversible degradation window.

[0011] Furthermore, the diffusion model generates candidate trajectories containing future virtual state sequences and retirement action markers, conditioned on the current state vector; the trajectory scoring function of the reinforcement learning integrates the discounted accumulation of future benefits, the penalty for falling into the irreversible degradation window, and the nonlinear penalty for retirement time delay.

[0012] Furthermore, during the reinforcement learning optimization process, the entropy of the generated trajectory distribution is calculated and added as a regularization term to the loss function to encourage policy exploration and prevent the model from getting trapped in local optima.

[0013] Furthermore, the behavior evaluation logic is implemented through a scoring function that integrates the predicted unit benefit at the time of retirement, the predicted performance degradation rate, and an indication of whether it falls into the irreversible degradation window, and determines the behavior type based on the scoring results.

[0014] Furthermore, S5 also includes: When applied to multi-module battery systems, a behavior score is calculated independently for each module, and a partial decommissioning recommendation list at the module level is output.

[0015] The beneficial technical effects of the present invention are at least as follows: To address the problem that existing methods cannot simultaneously characterize changes in the future use value of power batteries, potential irreversible degradation windows, and strategy selection in complex usage environments, this invention proposes an active decommissioning decision-making method based on time-domain state modeling, joint evaluation of benefits and degradation, and diffusion reinforcement learning trajectory generation.

[0016] By constructing a state vector that includes health status, internal resistance, temperature trend, load power, unit benefit prediction, and degradation rate, this invention can express the current state of the battery and the basis for its future evolution in a unified structure. By introducing a benefit-loss ratio sequence for future time periods and adding a modulation term for short-term benefit fluctuations, this invention can dynamically characterize the changing patterns of the battery's future use value. By combining thermal trends, impedance slope, voltage changes, and low-efficiency inversion signals to construct a risk scoring function, and predicting irreversible degradation time windows through deformation mapping, this invention achieves proactive identification of potential degradation critical intervals.

[0017] Based on this, the present invention uses a conditional diffusion model to generate multiple future strategy trajectories, and optimizes them through reinforcement learning using a trajectory score function that integrates cumulative benefits, risk penalties and nonlinear suppression of delayed retirement, thereby selecting the optimal retirement timing and strategy sequence.

[0018] Ultimately, this invention structures the retirement time and corresponding state information in the optimal trajectory into executable behavioral suggestions, achieving seamless implementation from model output to system scheduling. This invention enables proactive and forward-looking retirement strategy planning in complex and dynamic operating environments, effectively preventing batteries from entering irreversible damage zones while maximizing their usability. It is applicable to proactive retirement management of vehicle-mounted power batteries, energy storage systems, and multi-module battery packs. Attached Figure Description

[0019] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the temporal decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning, as proposed in this invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] In one or more embodiments, such as Figure 1 As shown, a time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning is disclosed. The method includes the following steps: S1. Construct a state vector for the predicted state; wherein, the state vector includes the health status, internal resistance, voltage, current, temperature trend, load power, unit benefit prediction, and performance degradation rate of the power battery.

[0023] Specifically, in order to generate a future-oriented active retirement strategy for power batteries, this step requires constructing a structured state vector ( The state vector is used to express the key operating status of the battery at the current moment, environmental characteristics, and predictive indicators needed for strategic decision-making. The first type of component of the state vector is the battery's operating status data. Health indicators ( The health value is obtained through the state estimation module built into the Battery Management System (BMS), which typically assesses changes in rated capacity based on cumulative charge and discharge capacity. In mainstream commercial energy storage systems, the BMS updates the health value every 5 minutes, and the system can read it directly via the CAN interface. Battery internal resistance ( This is achieved through AC impedance sampling. In practical applications, the AC perturbation method is used, for example, applying a small amplitude 0.1Hz AC signal to the battery cell and calculating the impedance value by measuring the corresponding voltage response. The system typically integrates an impedance measurement sub-board, which updates every 10 minutes and records the results to the system operation log.

[0024] Furthermore, the individual cell voltages of the battery were also collected simultaneously. ) and current ( Voltage is measured in real time by a sampling chip on the battery pack, with sampling accuracy at the millivolt level. Current sampling is achieved through a Hall effect sensor or a shunt resistor. In large battery systems, closed-loop Hall sensors are often used to improve dynamic response capabilities. To suppress the impact of sampling fluctuations on subsequent decisions, both data points use a 3-second time window moving average as input features.

[0025] Temperature index ( This is another indispensable state variable, collected by thermistors placed on the module casing or cell surface. The system reads the data through the thermal management unit and performs fusion processing. To enhance its value in decommissioning strategies, instantaneous temperature values ​​are not used here. Instead, the slope of the linear fit of the temperature change over the most recent 10 minutes is taken to construct a temperature trend index, which is used to characterize the growth trend of thermal runaway risk.

[0026] Furthermore, in addition to basic operational data, two predictive indicators need to be introduced to support future strategy decisions. One is the estimated unit usage benefit (…). This variable represents the economic benefit that each unit of energy output can generate over a future period. In the context of grid-connected energy storage, this variable is estimated using an electricity price prediction model. The model employs a three-layer fully connected neural network structure. The input consists of the electricity price sequence and load distribution information for the past 24 hours, with the structural dimensions set to [24, 16, 8, 1]. The activation function of the intermediate layer uses ReLU, and the output is the average unit electricity price for the next 5 minutes, which can be further converted into a unit energy benefit value. In the vehicle-mounted scenario, this variable is obtained through daily mileage, an energy consumption model, and mileage benefit parameters.

[0027] Another predictive metric is the rate of performance degradation ( The internal resistance is calculated through the time-series evolution trend. The system acquires data at each time (…). After that, it is combined with the historical impedance values ​​from the past 15 minutes to form a time series, and then a linear regression is performed. The slope is ( The value of ) is calculated as follows: ; in,( ) represents the internal resistance value at the current moment, ( ) represents the resistance value at the start of the time window, ( The ) indicates the length of the backtracking time, which is usually 15 minutes.

[0028] Furthermore, to ensure a consistent numerical scale, all variables in the state vector must be standardized before combination. Zero-mean normalization is employed, with each variable's mean and standard deviation calculated based on historical data and fixed during model training. The final state vector is defined as follows: ; in,( The current load power is collected by the power detection module of the power converter and continuously updated within a unit of time, with an accuracy that meets the requirements of the strategy judgment.

[0029] S2. Based on the unit benefit prediction and performance degradation rate, calculate the benefit-loss ratio sequence at each time point in the future, and identify the benefit window with the greatest future use value.

[0030] Specifically, in the decision-making scenario of proactively retiring power batteries, whether a battery should continue to serve cannot rely on a single health threshold or remaining lifespan judgment. Instead, it must be based on a quantitative analysis combining changes in economic benefits and performance degradation trends over a future period. This step utilizes the state vector formed in the previous stage. As input, the economic forecasts in the vector are used With the decline trend volume A benefit-loss ratio sequence is constructed for future time domains to evaluate the use value at different future moments and identify the benefit window with the highest cost-effectiveness. This process directly serves the core objective of the entire patent, enabling the subsequent strategy trajectory generation module to make decision biases based on a time-domain objective function.

[0031] First, set the future evaluation timeline. ,in The selection can be based on the application environment (such as power grid dispatch cycle or vehicle operation dispatch cycle). to Minutes. Unit efficiency at each future moment. Inference is achieved by the feedforward neural network prediction model defined in step one through sliding updates of the state input. The input includes historical load changes, price cyclical structure, and the current... and The output is an estimate of the economic benefit per unit energy at the corresponding time point. Future performance degradation rate. With the current Starting from the trend point, the value is obtained through linear extrapolation with a fixed step size. Based on the above predicted values, the basic benefit-loss ratio for each future time point is constructed, defined as follows: ; in, The predicted unit economic benefit comes from the inference output of the deployed three-layer neural network; The predicted performance degradation rate is obtained by extrapolation from the current linear fitting trend. For positive numbers, usually take This is used to avoid numerical instability caused by an excessively small denominator. From a physical perspective, Indicates a future moment The higher the ratio of the economic return that the system can obtain for each unit of performance degradation when it continues to use the battery, the more worthwhile it is to continue using the battery at that moment.

[0032] Batteries may experience fluctuations in efficiency or even brief peaks during their future service life. Therefore, to avoid misleading the strategy due to single-point instantaneous fluctuations, an additional stability adjustment term is introduced to penalize the short-term acceleration of predicted efficiency, thereby constructing a modulation ratio that better reflects the actual operating characteristics of power batteries. ; in, Stability adjustment parameters are set empirically based on system operating data; This is the second-order change estimated using the three-point difference method within the evaluation window, used to characterize the fluctuation trend of future benefit forecasts. Its function is to remove future "high-risk, high-volatility intervals" from the "high-benefit intervals," resulting in... It better represents the reliable and stable value range of power batteries in actual operation. For example, in scenarios where vehicle charging prices fluctuate drastically, even if there are short-term high economic benefits, if the benefits on both sides change rapidly, the second-order change term will increase significantly, leading to penalties. This avoids misjudging such "false high value points" as the continued use area before the optimal retirement time.

[0033] Furthermore, using the adjusted sequence Within the future window, a moving average and local extremum detection are performed at each time step, using a window length of [value missing]. The neighborhood interval is used to identify the set of moments within the interval that simultaneously satisfy the characteristics of local average maximum and fluctuation inflection (i.e., forward growth and backward decline), thereby forming a benefit window. This window represents the high-value segments most suitable for continued use in the near future. These segments will serve as the strategy bias direction in the subsequent trajectory generation stage, guiding the reward function during trajectory search based on time-domain characteristics.

[0034] The output of this step includes the modulation efficiency ratio sequence for each time step in the future interval. and the benefit window identified therefrom. Both will be directly input into the next stage's strategy trajectory generation module, participating as benefit items in the trajectory scoring process, thus enabling strategy generation to have a time-domain bias capability consistent with the actual economic behavior of the power battery.

[0035] S3. Based on the state vector and benefit-loss ratio sequence, and taking into account impedance growth, temperature change, voltage change, load and benefit factor, calculate the future risk score to predict the irreversible degradation window in which the battery may experience irreversible performance degradation.

[0036] In proactive battery retirement strategies, accurately identifying high-risk sections of the battery during future operation is a crucial component of the entire system's decision-making mechanism. The goal of this step is based on the current system state. and the benefit modulation sequence generated in the previous step Predicting the time window when the battery is about to enter an irreversible performance degradation state. Unlike traditional methods that only make a single-point static judgment on the current health status, this step is aimed at a future time domain. It combines the dynamic changes of thermal, load, and electrochemical behavior to construct a risk discrimination mechanism with a time evolution structure, thereby providing the input of boundary constraints and penalty functions for the trajectory generation stage.

[0037] Power batteries often experience "latent degradation" during use, meaning that performance indicators such as... and Even when the battery is still within the normal range, irreversible physical damage has already begun to accumulate, such as accelerated growth of the SEI film thickness, electrode pulverization, or the risk of internal short circuits. Therefore, the starting point of this step is to construct a future risk scoring structure to predict whether the battery, starting from its current state, will enter this irreversible process within a certain period of time.

[0038] First, in the future time domain interval For the analysis window, combine the data generated in step two. As a reference for the expected benefits of the current system, the risk of accelerated battery degradation due to "benefit-driven overuse" is considered. Therefore, a composite risk scoring function is designed. Its composition is as follows: ; in, This represents the impedance growth rate, which in this system is represented by... Obtained by forward linear extrapolation; It was measured by a temperature sensor. The slope of change over the past 5 minutes represents the trend of thermal runaway; The rate of voltage change is calculated from two adjacent voltage sampling points. The current load power is fed back in real time by the main controller; This is the modulation benefit sequence constructed in the previous step, used to reflect the expected value of the battery at that point in time. To prevent division by zero for extremely small numbers; all coefficients The scoring function is obtained by fitting a degenerate sample set during the system design phase, ensuring that it has the ability to distinguish in practice.

[0039] The core innovation of this scoring function lies in the introduction of a reverse value adjustment term, in addition to traditional degradation risk indicators (impedance, temperature, current). This is of strategic significance—the higher the benefit value, the more likely the model is to "continue using" the battery, while if objective indicators simultaneously show a deteriorating trend (such as...). If the increase is increased, the behavior may force the battery into the irreversible region. Therefore, the reverse adjustment term can effectively identify the high-risk area of ​​"risk-benefit conflict", thereby supplementing the risk of strategic misuse that cannot be reflected by traditional methods.

[0040] Based on the above scoring function, a standardized risk score is applied to each time step within the future interval. To enhance the model's ability to distinguish critical state transition points, a regularized deformation function is introduced. Mapped to Risk probability value within the interval : ; in, An adjustment factor to control the steepness of the mapping curve, This is the center offset constant, used to map "non-critical fluctuations" commonly seen in the system history to the neutral region. This function differs from the direct use of the standard sigmoid mapping by introducing an offset. It can effectively filter out the influence of "non-realistic decay fluctuations" and improve the model's sensitivity to critical points. In experiments, this structure showed better response performance than traditional thermal safety factor models, and is particularly suitable for potential failure identification in large-capacity, slowly changing scenarios.

[0041] Ultimately, the definition is not The reversible degradation time window is when the risk probability is greater than a set threshold. The set of time periods, namely: ; in The tolerance strategy can be adjusted according to different systems, and is usually set to... or Output results This will serve as a time boundary constraint for the trajectory strategy generation phase, preventing the strategy path from being delayed until the battery enters the irreversible degradation phase before the retirement action is executed; and As a continuous risk distribution, it can be used as a penalty term in the trajectory path scoring function to dynamically adjust the trajectory evaluation score and form a model structure for comprehensive risk-benefit evaluation.

[0042] S4. Using the state vector as a condition, generate multiple future policy trajectories containing future states and retirement actions using a diffusion model, and use reinforcement learning combined with the benefit-loss ratio sequence and the irreversible degradation window to score and optimize the trajectories, and select the optimal retirement trajectory.

[0043] The generation of the strategy trajectory is a key step in this invention, connecting the analysis module and the behavior output. Its core task is to generate the strategy trajectory given the current state of the system. Future benefit assessment sequence With irreversible degradation window Given the known conditions, we dynamically plan an optimal retirement path. Unlike traditional reinforcement learning methods that directly select actions based on the policy network, this scenario involves generating the policy space at the future point in time, which has temporal extensibility and structural complexity. Therefore, this step constructs a trajectory generation structure based on a conditional diffusion model and introduces a reinforcement learning module to score and optimize the trajectory, realizing a complete learning loop from "generation → scoring → updating".

[0044] The trajectory generation model is constructed based on a conditional diffusion process. The input is the current system state. The variables it contains come from the sensor acquisition and real-time system state estimation module in step one: Obtained from the capacity estimation module within the battery management system. AC disturbance test results from the impedance measurement sub-board. and Data is collected in real time using a voltage detection chip and a current Hall sensor, respectively. For the output of the module's environmental thermistor, Read the current load power from the Energy Management System (EMS) interface. and These originate from the economic forecasting neural network and the impedance change slope fitting module constructed in step one, respectively. The trajectory generation interval is... , It is generally set at 30 minutes, divided into... Each discrete time step.

[0045] During the training phase, the diffusion model uses historical real trajectories as the target, mapping them to a progressively noisy latent representation. The model structure employs a residual diffusion network, with each diffusion step consisting of three modules: the input is the latent trajectory state of the current step; after adding temporal encoding, it enters a one-dimensional convolutional layer to extract temporal features; and then it is fused with the information from the previous layer through skip connections. In the reverse path of the diffusion process, the current system state is introduced. As a condition, future trajectories are generated by reconstructing the conditional residual blocks. Each trajectory is in the form of... Each of them This represents the virtual state of the system at that moment and the retirement action flag (1 indicates retirement, 0 indicates continued operation). Trajectory generation does not directly predict actions, but rather generates a "future time-domain behavior sequence", where the first moment where action 1 occurs is the candidate retirement time point.

[0046] To quantify the merits of each trajectory, a trajectory scoring function was designed that combines benefit maximization, risk aversion, and time delay penalty. The trajectory score is first weighted and accumulated at each time point, with the base gain term being... This refers to the modulation benefit ratio sequence from step two, which is discounted and summarized over time in the trajectory to reflect the temporal structure of future benefits. This is combined with the degradation window predicted in step three. To prevent the strategy from entering high-risk periods, a risk penalty term is added, and a time deformation function is introduced as a delay penalty structure to mitigate and suppress the resource waste caused by premature retirement. The final trajectory score is defined as follows: ; in, This is the time discount factor, used to control the impact of future returns. It is an indicator function for the retirement time in the trajectory. fall into The value is 1 at time. This indicates the severity of the penalty for entering a high-risk area; and The penalty slope and inflection point for retirement delay are controlled separately to ensure that the retirement time is not too early, thus preventing potential benefit loss. The innovation of this trajectory scoring mechanism lies in that it not only dynamically integrates future benefits and system physical state, but also introduces a "smooth policy boundary" through a nonlinear deformation function, avoiding policy rigidity caused by hard thresholds.

[0047] Furthermore, in the policy optimization phase, the reinforcement learning module uses trajectory scores... The trajectory generation model is updated using policy gradients. Considering that the output of the diffusion generator is essentially a conditional distribution generation process, the model is treated as an implicit policy network, and its output represents the current state. Trajectory distribution under certain conditions , in This represents the learnable parameters in the diffusion model. The trajectory generated at each sampling point. Assigned a corresponding score This serves as the evaluation signal for the policy instance. The policy gradient estimation process follows the REINFORCE algorithm framework, estimating the gradient of the sampled trajectory in the following desired form: ; The actual calculation of this gradient term involves sampling multiple trajectories. And calculate its score. Implementation. During implementation, to prevent the model from falling into mode collapse or policy degradation, distribution entropy is introduced as a regularization term to encourage exploration of the trajectory space. Specifically, in each round of parameter updates, the Shannon entropy of the current model's output trajectory distribution is calculated: ; The entropy term is added as a regularization term in the loss function to the objective function, forming the final optimization objective. : ; in The weights of the entropy regularization term are hyperparameters, typically determined based on the performance on the validation set. Adjustments are made between these parameters. The optimization objective is solved iteratively using stochastic gradient descent or the Adam optimizer. The model's training data consists of historical running data, including the state... The system builds a real-time monitoring dataset, and historical trajectories are derived from strategy replays in decommissioned label samples or high-dimensional operation logs.

[0048] The final output is the optimal trajectory. The trajectory contains a complete state-action sequence from the current state to retirement, and also provides a clear recommended retirement time. This refers to the time point when the action flag in the trajectory first appears as 1. This result will serve as the basis input for the final policy behavior, proceeding to the next step to generate control recommendations for the battery management system or operation scheduling system.

[0049] S5. Based on the recommended retirement time and its corresponding status determined by the optimal retirement trajectory, and combined with the preset behavior evaluation logic, generate executable behavior suggestions including immediate retirement, delayed observation, or continued operation.

[0050] Specifically, the core task of this step is to obtain the optimal strategy trajectory. With recommended retirement time Based on this, it outputs behavioral suggestions that can be directly executed by the system management, achieving a closed loop from model decision-making to engineering implementation. Input information comes from the previous stage's strategy trajectory generation module, specifically including complete trajectory sequences. Each time step The system's operational status at that moment, whether it has been decommissioned, and the time point at which the decommissioning action first occurred. In addition, the irreversible degradation window predicted in step three also needs to be used. This is used to assist in determining whether the system risk boundary has been reached.

[0051] First, it is necessary to extract from the trajectory Corresponding system status data This state is determined by the set of variables defined in step one at time. The derived values ​​are composed of the following, namely .in and These represent the battery health and internal resistance prediction values, respectively, derived from the current battery state inference model. The result obtained by combining the trajectory generator recursion; and These are the predicted values ​​of the voltage and current trajectories, respectively, and are the outputs from the residual diffusion network during the trajectory generation stage. The time series modeling results are derived from the module temperature sensor data, which is collected every 30 seconds using thermistors distributed inside the battery module. It is the predicted load signal fed back from the power converter; It is the unit benefit value inferred by the feedforward neural network model based on electricity price forecasts and load history, while The future impedance growth rate is obtained by extrapolation from a linear fitting model.

[0052] After obtaining After collecting all system status data, the system will execute behavioral evaluation logic to determine the type of decommissioning operation. To this end, a behavioral scoring function is designed for this step. The function uses a linear combination of key variables to determine whether to immediately execute the decommissioning order. The function structure is as follows: ; In this function, Indicates unit economic benefits, Indicates the predicted rate of performance degradation; For indicator functions, if The value is 1 if the value falls into the irreversible degradation window; coefficient , and These are configurable parameters that can be set based on the user's risk appetite and sensitivity to economic returns. For example, for energy storage scenarios where safety is paramount, the following settings can be adopted: , For electric mobility systems such as taxis, settings can be configured. The ratio is relatively high, prioritizing maximizing the unit's pre-retirement benefits.

[0053] Furthermore, based on the result of this function, the system will automatically output the behavior type. If This indicates that the cost of decommissioning exceeds its potential value, and decommissioning should be carried out immediately; if (in For buffer parameters determined empirically (e.g., 0.05), the system recommends entering the "observation period" state, which maintains the current operating state but sets alarm thresholds for dynamic observation; if If the current operating state of the system still has continuing value, the retirement decision will be automatically postponed.

[0054] To enhance the system's operability under complex structures, this step supports module-level strategy decomposition. If the trajectory... The middle corresponds to a multi-module battery structure, each module Each has its own independent state sequence and recommended retirement time. Then the behavior function The decommissioning will be performed in parallel, and this will be used to output a module-level decommissioning recommendation table. For example, if in a 10-module system, only 3 modules enter the degradation window during prediction and have a negative decommissioning score, the system will automatically output a "partial decommissioning" recommendation, and add fields such as a list of module numbers, operation time limits, and dismantling instructions to the behavior recommendation structure.

[0055] The final output behavior suggestion structure will be encapsulated into the system instruction interface in the following format: Recommended retirement time: ; Behavior type: Immediate decommissioning / Delayed observation / Continuous operation; List of module numbers (if applicable); Additional suggestions: such as whether temperature anomalies trigger adjustments to the thermal management system controller in parallel; Control flags: written to the BMS control domain for use in issuing CAN commands.

[0056] This output structure can be directly written into the main control system controller and transmitted to the battery management system (BMS), energy management system (EMS), or vehicle control system (VCU) via CAN bus or MODBUS protocol, achieving seamless integration from strategy recommendations to system operation.

[0057] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0058] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0059] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0060] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0061] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0062] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0063] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A temporal decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning, characterized in that, Includes the following steps: S1. Construct a state vector for the predicted state; wherein, the state vector includes the health status, internal resistance, voltage, current, temperature trend, load power, unit benefit prediction, and performance degradation rate of the power battery; S2. Based on the unit benefit prediction and performance degradation rate, calculate the benefit-loss ratio sequence at each time point in the future, and identify the benefit window with the greatest future use value. S3. Based on the state vector and benefit-loss ratio sequence, and taking into account impedance growth, temperature change, voltage change, load and benefit factor, calculate the future risk score to predict the irreversible degradation window in which the battery may experience irreversible performance degradation. S4. Using the state vector as a condition, generate multiple future policy trajectories containing future states and retirement actions using a diffusion model, and use reinforcement learning combined with the benefit-loss ratio sequence and the irreversible degradation window to score and optimize the trajectories, and select the optimal retirement trajectory. S5. Based on the recommended retirement time and its corresponding status determined by the optimal retirement trajectory, and combined with the preset behavior evaluation logic, generate executable behavior suggestions including immediate retirement, delayed observation, or continued operation.

2. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, The performance degradation rate is calculated by the time-series evolution trend of the internal resistance.

3. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, The benefit-loss ratio sequence for each time point in the future is generated by calculating the unit benefit and performance degradation rate prediction value for each time point.

4. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 3, characterized in that, When calculating the benefit-loss ratio at each moment in the future period, a stability adjustment term for the second-order change of the unit benefit prediction value is introduced to correct the error. This is used to suppress the misleading effect of short-term high volatility intervals on decision-making, and finally obtain the benefit-loss ratio sequence at each moment in the future period.

5. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, The benefit factor is calculated based on the benefit-loss ratio sequence at each time point within the future period.

6. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, S3 further includes: A standardized risk score is calculated for each time point within the future interval, and this future risk score is mapped to... Risk probability value within the range; The set of time periods in which the risk probability value is greater than a preset threshold is used as the irreversible degradation window.

7. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, The diffusion model generates candidate trajectories containing future virtual state sequences and retirement action markers, conditioned on the current state vector; the trajectory scoring function of the reinforcement learning integrates the discounted accumulation of future benefits, the penalty of falling into the irreversible degradation window, and the nonlinear penalty of retirement time delay.

8. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 7, characterized in that, In the reinforcement learning optimization process, the entropy of the generated trajectory distribution is calculated and added as a regularization term to the loss function to encourage policy exploration and prevent the model from getting stuck in local optima.

9. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 1, characterized in that, The behavior evaluation logic is implemented through a scoring function that integrates the predicted unit benefit at the time of retirement, the predicted performance degradation rate, and an indication of whether it falls into the irreversible degradation window, and determines the behavior type based on the scoring results.

10. The time-domain decision generation method for active decommissioning of power batteries based on diffusion reinforcement learning according to claim 9, characterized in that, S5 also includes: When applied to multi-module battery systems, a behavior score is calculated independently for each module, and a partial decommissioning recommendation list at the module level is output.

Images