A regenerative braking energy storage scheduling method and system based on dynamic prediction

Abstract

This invention discloses a method and system for regenerative braking energy storage scheduling based on dynamic prediction. The method includes: collecting data on vehicle operation, line conditions, environment, energy storage status, grid load, and real-time carbon price; generating predicted values ​​for train braking energy, multi-vehicle conflict probability, and energy storage health status through a prediction model, and calculating expected carbon emission reduction benefits; inputting the above results into a decision model, which uses trains, energy storage systems, the grid, and carbon trading nodes as game players, and performs collaborative optimization through multi-agent reinforcement learning fusion game strategies, outputting charging and discharging scheduling commands; issuing execution commands through a cloud-edge-device architecture, with edge nodes autonomously responding to emergencies in case of communication interruption. The system includes corresponding modules. This invention achieves proactive collaborative optimization from passive storage to "prediction-game theory-maintenance-carbon value-added," improving energy utilization and economic efficiency.

Description

Technical Field

[0001] This invention relates to the field of regenerative braking energy management and energy storage scheduling technology for rail transit, and in particular to a method and system for regenerative braking energy storage scheduling based on dynamic prediction. Background Technology

[0002] With the continuous expansion of rail transit networks, the regenerative energy generated during train braking has become an important source for improving system energy efficiency. To achieve effective recovery and reuse of this type of energy, the industry generally adopts energy dispatch strategies based on energy storage devices. However, current dispatch methods are still mainly based on preset thresholds or simple rules, and usually perform coarse charging and discharging control based on the state of charge of energy storage devices, the approximate location of trains, or grid load, lacking dynamic adaptability to actual operating scenarios.

[0003] On lines with multiple trains operating intensively, the braking periods of trains overlap, resulting in a significant increase in the instantaneous peak input of regenerative energy. Existing scheduling strategies often cannot identify or predict such conflicts in advance and can only respond passively based on single events. This leads to uneven distribution of energy storage resources when multiple trains compete for resources, resulting in insufficient recovery of regenerative energy for some high-demand trains and thus energy waste.

[0004] Furthermore, rail transit systems are typically located in complex environments characterized by high humidity, high temperature, and low temperature. Environmental factors have a significant impact on the lifespan and performance of energy storage systems. However, most existing energy storage health prediction methods do not incorporate environmental conditions into their modeling, leading to significant biases in health status assessments and hindering the long-term stable operation of energy storage systems.

[0005] Meanwhile, in addition to its electrical value, regenerative braking energy also possesses potential carbon emission reduction value. With the gradual promotion of carbon trading mechanisms, linking renewable energy recovery with carbon market revenue has become a key focus for the industry. However, most existing dispatch methods only aim at energy utilization efficiency, neglecting the carbon emission reduction benefit dimension, resulting in the failure to maximize the overall value of the system.

[0006] In terms of system architecture, with the development of the "cloud-edge-device" collaborative operation mode, scheduling logic increasingly relies on cloud models for global decision-making. However, in tunnels or underground sections, communication links are easily interrupted by environmental factors, making it difficult for edge nodes to obtain the latest scheduling instructions in a timely manner. Existing edge devices can usually only execute fixed strategies and lack the ability to autonomously generate emergency scheduling, affecting the reliability and security of the system.

[0007] Based on the above problems, there is an urgent need for a regenerative braking energy dispatching method that can integrate operational status prediction, energy storage health assessment, conflict prediction, carbon value calculation, and chain break emergency response capabilities, in order to achieve multi-objective collaborative optimization and improve the overall efficiency of energy storage systems and train operation. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for scheduling regenerative braking energy storage based on dynamic prediction, which can achieve multi-objective collaborative optimization and improve the overall efficiency of energy storage systems and train operation.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a regenerative braking energy storage scheduling method based on dynamic prediction, applied to a collaborative scheduling system including a cloud platform, edge nodes, and on-board and trackside execution terminals, the method comprising:

[0010] Step S1: Collect vehicle operation data through sensors deployed on the train, collect track condition data and environmental status data through sensors deployed along the track, collect energy storage status data through the battery management system of the energy storage cabinet, and obtain grid load data and real-time carbon price data through the data interface.

[0011] Step S2: Input the vehicle operation data, the line condition data, and the power grid load data into a short-term prediction model to generate a predicted value for train braking energy and the probability of multi-vehicle energy conflict for a future preset period; input the energy storage status data and the environmental status data into a lifetime prediction model to generate a predicted value for the health status of the energy storage system; and calculate the expected carbon emission reduction benefits based on the predicted value for train braking energy and the real-time carbon price data.

[0012] Step S3: Input the predicted value of the train braking energy, the probability of energy conflict among multiple vehicles, the predicted value of the health status, and the expected carbon emission reduction benefits into a decision model; The decision model takes the train, energy storage system, power grid, and carbon trading node as game participants, integrates game theory strategies through a multi-agent reinforcement learning algorithm, coordinates and optimizes the comprehensive benefits of each participant, and outputs charging and discharging scheduling instructions for the energy storage system.

[0013] Step S4: Through the cloud-edge-device architecture of the collaborative scheduling system, the charging and discharging scheduling command is sent to the corresponding energy storage system for execution; wherein, when the edge node detects a communication interruption with the cloud platform, it autonomously generates and sends out an emergency scheduling command based on the locally cached prediction model and decision logic.

[0014] Furthermore, in step S2, the short-term prediction model is a deep learning model built on a long short-term memory network, and its input feature dimension is N-dimensional, including real-time speed and acceleration in the vehicle operation data, slope and curvature in the road condition data, and historical braking energy sequence, where N is an integer greater than 1.

[0015] The deep learning model is used to output the predicted value of the train braking energy and the probability of multi-vehicle energy conflict within a first preset time period in a rolling prediction manner; wherein, the predicted value of the train braking energy Calculated from a regression subnetwork in the deep learning model:

[0016] ;

[0017] in, Let be the error function. It is a zero-order Bessel function. It is the natural logarithm function. This is a calibration coefficient for the system's energy conversion efficiency, used to calibrate the overall energy conversion efficiency of the formula. The number of trains participating in the prediction. For the first The train's braking force request characteristic coefficients are used to map real-time acceleration to weights for braking force requests. For the first Real-time acceleration data collected by the train's acceleration sensors. For the first The train speed fluctuation weighting factor is used to control the scaling of the difference between the real-time speed and the reference speed on the Bessel function input. For the first The difference between the train's real-time speed and the reference speed. For the current moment, The duration of the first preset duration, This is a time-varying power density function reconstructed based on real-time speed and slope data, used to characterize the power density at time t. The instantaneous power change rate, The integral variable in this formula is [variable name]; the range of the predicted train braking energy value is [range]. The unit is kilowatt-hour (kWh);

[0018] The probability of multi-vehicle energy conflict is calculated by a classification subnetwork in the deep learning model based on the same input features.

[0019] Further, in step S2, the lifetime prediction model is an electrochemical-environmental coupled degradation model; the electrochemical-environmental coupled degradation model is used to dynamically correct the lithium-ion diffusion coefficient and interface impedance growth rate inside the battery based on the battery state of charge, health state, and temperature in the energy storage state data, and the humidity and electromagnetic intensity in the environmental state data, and to predict the health state of the energy storage system based on the corrected parameters. :

[0020] ;

[0021] in, It is an exponential function. This represents the nominal value of the initial health state of the energy storage system. The cumulative operating time of the energy storage system. The temperature data is from the environmental condition data. The humidity data is from the environmental status data. The temperature, humidity, and stress coupling function is defined as follows: ,in and The material aging response coefficient. For reference to ambient temperature, For reference to ambient temperature and humidity, It is the natural logarithm function. The charging and discharging current data in the energy storage state data. The electromagnetic intensity data is from the environmental state data. The electrochemical-electromagnetic stress function is defined as follows: ,in The electrochemical reaction rate constant is Electromagnetic interference factor, Let be the error function. This is the reference value for electromagnetic intensity. The integral variable in this formula; the predicted health status value The range of is (0,1], and it is a dimensionless ratio.

[0022] Furthermore, in step S2, the expected carbon emission reduction benefit is calculated as follows: the predicted value of train braking energy is converted into expected recyclable electrical energy, and combined with the real-time carbon price data and the real-time carbon quota information obtained from the carbon trading market, a calculation model that includes a carbon price elasticity factor and time validity is used for dynamic calculation.

[0023] Furthermore, in step S3, the decision model is a hybrid decision model constructed based on the multi-agent deep deterministic policy gradient algorithm and the game tree search algorithm;

[0024] The hybrid decision-making model includes a central evaluation network and multiple executor networks corresponding to different game participants. The central evaluation network is used to output a comprehensive reward value based on global state information and the payoff function of each game participant. Each executor network is used to optimize the strategy based on local observation information and the comprehensive reward value and output the action strategy. Finally, the charging and discharging scheduling instruction is generated based on the game equilibrium result of each action strategy.

[0025] Furthermore, the payoff function of the train as a game participant integrates instantaneous energy recovery efficiency and real-time operational safety indicators, while the payoff function of the energy storage system as a game participant integrates the lifespan attrition rate determined based on the predicted health status value and the economic benefits including the expected carbon emission reduction benefits.

[0026] Furthermore, in step S3, when the decision model is running, it dynamically adjusts the weights of each game participant in the calculation of the comprehensive reward value based on at least one of the following conditions: whether the grid load exceeds the first load threshold, whether the predicted health status value is lower than the first health threshold, or whether the probability of multi-vehicle energy conflict is higher than the first conflict threshold.

[0027] Furthermore, in step S4, the cloud-edge-device architecture uses a dual-mode redundant communication link composed of 5G communication and Beidou short message communication for data transmission; when the edge node detects that the communication delay with the cloud platform exceeds the first delay threshold and the signal strength is lower than the first strength threshold, it determines that the communication is interrupted and triggers the local generation of the emergency dispatch command.

[0028] Furthermore, the prediction model locally cached at the edge node is a lightweight version of the short-term prediction model and the lifetime prediction model obtained through model pruning and quantization techniques; the generation logic of the emergency dispatch command prioritizes ensuring the safe operation of the train's emergency braking energy recovery and energy storage system.

[0029] A regenerative braking energy storage scheduling system based on dynamic prediction, applied to the regenerative braking energy storage scheduling method based on dynamic prediction as described above, includes:

[0030] The data acquisition module is used to collect vehicle operation data, line condition data, environmental status data, energy storage status data, power grid load data, and real-time carbon price data.

[0031] The collaborative prediction module, connected to the data acquisition module, is used to input the vehicle operation data, the line condition data, and the power grid load data into a short-term prediction model to generate a predicted value for train braking energy and the probability of multi-vehicle energy conflict for a future preset period; input the energy storage status data and the environmental status data into a lifetime prediction model to generate a predicted value for the health status of the energy storage system; and calculate the expected carbon emission reduction benefits based on the predicted value for train braking energy and the real-time carbon price data.

[0032] A hybrid decision-making module, connected to the collaborative prediction module, is used to input the predicted value of train braking energy, the probability of multi-vehicle energy conflict, the predicted value of health status, and the expected carbon emission reduction benefits into a decision model. The decision model takes the train, energy storage system, power grid, and carbon trading node as game participants, integrates game theory strategies through a multi-agent reinforcement learning algorithm, collaboratively optimizes the comprehensive benefits of each participant, and outputs charging and discharging scheduling instructions for the energy storage system.

[0033] The instruction execution and emergency module is connected to the hybrid decision-making module and the data acquisition module, respectively. Through the cloud-edge-device architecture of the collaborative scheduling system, the charging and discharging scheduling instructions are sent to the corresponding energy storage system for execution. When the edge node detects a communication interruption with the cloud platform, it autonomously generates and sends out emergency scheduling instructions based on the locally cached prediction model and decision logic.

[0034] The beneficial effects of this invention are:

[0035] The regenerative braking energy storage scheduling method based on dynamic prediction provided by this invention can overcome the shortcomings of existing technologies in terms of insufficient conflict prediction, weak environmental adaptability, single value dimension, and limited cloud-edge collaboration capabilities. Specifically, it has the following beneficial effects:

[0036] 1. Achieve early conflict avoidance in multi-train braking scenarios and improve energy recovery efficiency: By inputting vehicle operation data, line condition data, and grid load data into a short-term prediction model, the predicted value of train regenerative braking energy can be obtained in advance within a preset future time period, and the probability of conflict due to overlapping energy inputs from multiple trains can be identified. Based on this, the decision-making model can optimize allocation before conflicts occur, avoiding resource bottlenecks caused by passive "first-come, first-served" allocation, improving energy access capacity in multi-train parallel braking scenarios, and significantly reducing energy waste.

[0037] 2. Integrating environmental conditions for energy storage health prediction improves the accuracy and safety of energy storage system lifespan assessment: This invention inputs actual environmental condition data and energy storage condition data into the lifespan prediction model, which can dynamically reflect the performance changes of the energy storage system in complex environments such as high humidity and low temperature, making the health condition prediction more realistic, reducing lifespan assessment errors, helping to reduce the risk of rapid degradation induced by the environment, thereby extending the service life of the energy storage unit and improving the safety of system operation.

[0038] 3. Introducing carbon emission reduction value calculation to achieve synergistic optimization of energy utilization and economic benefits: Based on predicted regenerative braking energy and real-time carbon price data, this invention calculates the obtainable carbon emission reduction benefits and inputs them as one of the decision factors into the multi-agent decision model, so that the scheduling strategy not only focuses on the amount of energy recovered, but also takes into account the carbon trading benefits, thereby achieving multi-objective optimization and improving the overall system value.

[0039] 4. Achieving the optimal solution for multi-node collaborative scheduling through multi-agent reinforcement learning fusion game strategy: The decision model incorporates trains, energy storage systems, power grids and carbon trading nodes into a unified framework. Through multi-agent reinforcement learning fusion game strategy, the scheduling decision can seek the comprehensive optimal solution among multiple stakeholders, achieving a comprehensive improvement in energy utilization efficiency, energy storage health and carbon revenue, which is significantly better than the static scheduling method based on simple rules.

[0040] 5. Possesses emergency response capability for cloud-edge disconnection, improving the continuity and reliability of system operation: When communication between edge nodes and the cloud platform is interrupted, edge nodes can autonomously generate emergency dispatch instructions based on locally cached models, no longer relying on fixed preset strategies, thereby ensuring the stable operation of the system in complex communication environments such as tunnels and underground sections, and improving the online rate and robustness of the dispatch system.

[0041] In summary, this invention significantly improves the recovery efficiency of regenerative braking energy, the health of energy storage systems, and the economic and carbon benefit value of the entire transportation system through the synergistic design of predictive driving, environmental integration, multi-objective optimization, and chain-breaking autonomous capabilities, demonstrating significant technical effects and application potential. Attached Figure Description

[0042] Figure 1 This is a flowchart of the steps of the regenerative braking energy storage scheduling method based on dynamic prediction in this invention;

[0043] Figure 2 This is a schematic diagram of the regenerative braking energy storage scheduling system based on dynamic prediction in this invention.

[0044] Attached reference numerals: 1. Data acquisition module; 2. Collaborative prediction module; 3. Hybrid decision-making module; 4. Command execution and emergency response module. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0046] Example 1, refer to Figures 1 to 2This is the first embodiment of the present invention. This embodiment provides a regenerative braking energy storage scheduling method based on dynamic prediction, applied to a collaborative scheduling system including a cloud platform, edge nodes, and on-board and trackside execution terminals. This method achieves efficient recovery and energy storage scheduling of train regenerative braking energy through a closed-loop mechanism of "prediction-evaluation-decision-execution".

[0047] Working principle of Example 1:

[0048] In step S1, comprehensive data acquisition and aggregation are performed. Sensors deployed on the train (such as speed sensors and accelerometers) collect real-time vehicle operation data, including key dynamic information such as speed and acceleration. A sensor network deployed along the track collects track condition data (such as gradient and curvature) and environmental condition data (such as ambient temperature, humidity, and electromagnetic field strength). The battery management system (BMS) of each energy storage unit continuously reports its energy storage status data, including battery state of charge (SOC), state of health (SOH), temperature, and charging / discharging current. In addition, the system obtains grid load data from the grid operator and real-time carbon price data from the carbon trading market through data interfaces. This multi-source heterogeneous data forms the data foundation for subsequent intelligent prediction and decision-making.

[0049] In step S2, the system uses two types of prediction models to perform forward calculations on critical states.

[0050] On one hand, vehicle operation data (real-time speed, acceleration), track condition data (gradient, curvature), and power grid load data (which may implicitly affect the train's traction / braking strategy) are input into a short-term prediction model. This model is a deep learning model built on a Long Short-Term Memory (LSTM) network. Its input feature dimension is N (N>1), and in addition to the aforementioned real-time data, it also includes historical braking energy sequences, enabling the model to learn the temporal pattern of energy generation. The model outputs two key indicators within a first preset time period (e.g., the next 30 seconds) using a rolling prediction method: the predicted value of train braking energy. Probability of energy conflict with multiple vehicles. Predicted train braking energy. It is calculated from a regression subnetwork, and its calculation formula is as follows:

[0051] ;

[0052] in Let be the error function. It is a zero-order Bessel function. It is the natural logarithm function. This is a calibration coefficient for the system's energy conversion efficiency, used to calibrate the overall energy conversion efficiency of the formula. The number of trains participating in the prediction. For the first The train's braking force request characteristic coefficients are used to map real-time acceleration to weights for braking force requests. For the first Real-time acceleration data collected by the train's acceleration sensors. For the first The train speed fluctuation weighting factor is used to control the scaling of the difference between the real-time speed and the reference speed on the Bessel function input. For the first The difference between the train's real-time speed and the reference speed. For the current moment, The duration is the first preset duration. This is a time-varying power density function reconstructed based on real-time speed and slope data, used to characterize the power density at time t. The instantaneous power change rate, The integral variable in this formula; the range of the predicted train braking energy value is... The unit is kilowatt-hour (kWh).

[0053] Error function Used to handle acceleration (Negative values ​​represent braking) nonlinear mapping. The S-shaped characteristic of the error function can simulate the process of braking demand from zero to saturation, thus more accurately reflecting the marginal effect of braking force demand on energy contribution;

[0054] Zero-order Bessel function Used to depict speed fluctuations Impact on the stability of braking energy recovery. The Bessel function has oscillatory decay characteristics, which can effectively simulate the suppression effect of large speed fluctuations on energy recovery efficiency, and improve the accuracy of the predicted value under non-stationary operating conditions.

[0055] Integral term Time-varying power density function reconstructed from real-time speed and slope data Calculate the total power integral over a future time period. Outer layer application. The function ensures that the result is positive (consistent with the energy properties) and also plays a certain role in smoothing out extreme high-power scenarios, preventing the predicted value from being distorted due to instantaneous spikes, thus enhancing the robustness of the model.

[0056] This formula integrates multi-vehicle information, nonlinear braking characteristics, speed fluctuation effects, and time-varying power of the track, ultimately outputting a predicted value for train braking energy. (Unit: kWh) provides high-precision energy input prediction for subsequent scheduling decisions, directly solving the problem of lagging or ineffective scheduling of energy storage systems caused by the coarse prediction of traditional methods.

[0057] Probability of energy conflict among multiple vehicles It is calculated by a classification sub-network within the same deep learning model based on the same input features. The specific calculation process is as follows: the classification sub-network first outputs a multi-dimensional conflict feature vector. This vector characterizes the potential risk intensity of overlapping braking energy releases from different trains in time and space. Then, it is processed by a probabilistic function... This is mapped to scalar probability values. The formula for the probabilistic function is:

[0058] ;

[0059] in express Activation function The weights of each conflict feature, For bias terms, The first of the conflict feature vectors Each component. The activation function compresses the weighted sum to the (0,1) interval, which is the probability of multi-vehicle energy conflict. This probability quantifies the risk that, within a predetermined future timeframe, the braking events of multiple trains may occur too close together, potentially causing the total braking power to exceed the maximum capacity of the associated energy storage system or triggering severe fluctuations in the DC grid voltage. Its technical solution lies in providing a crucial collaborative constraint signal for the decision-making model, enabling scheduling strategies to anticipate and mitigate energy congestion risks, thereby ensuring the stability of the traction power supply system and the safety of energy storage equipment.

[0060] On the other hand, energy storage state data (battery SOC, SOH, temperature, current) and environmental state data (humidity, electromagnetic field strength) are input into a lifetime prediction model. This model is an electrochemical-environment coupled degradation model, the core of which lies in dynamically correcting the internal electrochemical parameters of the battery to reflect the combined effects of complex operating conditions and environmental stress. This model predicts the health status of the energy storage system. The calculation formula is:

[0061] ;

[0062] in, This is the nominal value of the initial health state of the energy storage system, serving as the benchmark for attenuation calculation. This represents the nominal value of the initial health state of the energy storage system. The cumulative operating time of the energy storage system. Temperature data is from the environmental condition data. This refers to humidity data within the environmental condition data.

[0063] The temperature and humidity stress coupling function is defined as follows:

[0064] This function quantifies the coupling acceleration effect of ambient temperature and humidity on the battery aging rate. (Square term) It emphasizes the nonlinear acceleration of aging rate by temperature deviation from the reference value (especially high temperature) (which usually conforms to the spirit of the Arrhenius equation). For reference to ambient temperature, For reference to ambient temperature and humidity, It is the natural logarithm function. This refers to the charging and discharging current data in the energy storage status data. The terms describe the effect of humidity changes on internal side reactions (such as corrosion) of the battery, and the logarithmic form reflects the sensitivity of humidity effects within a certain range. and This represents the material aging response coefficient.

[0065] The electrochemical-electromagnetic stress function is defined as follows:

[0066] This function quantifies the effects of operating current and electromagnetic environment on aging. The electrochemical reaction rate constant is This partially reflects the quadratic relationship between charge / discharge current (especially high current) and battery life through Joule heating and electrochemical polarization. The term introduces electromagnetic intensity The impact, This is the reference value for electromagnetic intensity. Let be the integral variable in this formula. Error function. Electromagnetic interference factor The scope of its function is limited to The range was used to simulate the potential catalytic effect of electromagnetic fields on the growth of interfacial impedance (when...). Much larger than the benchmark value (The time tends to saturate).

[0067] Integral term The stress function described above will be applied over the cumulative battery operating time. Integrating the components essentially adds up the cumulative "damage" the battery experiences throughout its service life. (Outer layer) The function maps cumulative damage to an exponential decay of health status, which is consistent with many battery aging mechanism models.

[0068] This model achieves dynamic and refined prediction of battery health status by coupling environmental stresses (temperature, humidity, electromagnetic fields) with operational stresses (current). It addresses the problem of traditional methods that rely solely on cycle counts or simple ampere-hour integrals for lifetime estimation, neglecting the influence of complex environmental and operating conditions. The output health status prediction value is shown below. (range This provides the decision-making model with key lifespan status information of the energy storage system, enabling the scheduling strategy to proactively avoid stress impacts on aging batteries and extend the overall lifespan of the system.

[0069] Furthermore, in step S2, the system is based on the predicted value of train braking energy. Using real-time carbon price data and real-time carbon allowance information, a calculation model dynamically calculates the expected carbon emission reduction benefits. The specific calculation method is as follows: First, the predicted value of train braking energy is... Multiply by a local power grid carbon emission factor (Unit: kgCO2 / kWh) This factor is obtained from real-time carbon quota information to determine the expected carbon emission reduction. Then, using a factor incorporating carbon price elasticity... With time effectiveness factor The computational model performs dynamic calculations. The specific computational model formula is as follows:

[0070] .

[0071] in, express Real-time carbon price data at any given moment.

[0072] Carbon price elasticity factor (Positive or negative) Used to adjust for short-term fluctuations in carbon prices (Current carbon price and recent average carbon price) The impact of the difference () on earnings assessment. This simulates the market's reaction to price trends, assuming carbon prices are in a rapid upward trend ( >0 and If the value is greater than 0, then the future returns should be overestimated to incentivize the system to store energy more actively; otherwise, a conservative assessment should be made.

[0073] Time validity factor, defined as ,in The attenuation coefficient is... This represents the predicted length of the future period. This factor reflects the present value discount of future earnings. The larger the value, the smaller the factor, which means that the carbon emission reduction benefits in the more distant future are discounted. This is in line with the time value principle in economic decision-making and makes the benefit calculation closer to the actual financial logic.

[0074] This indicates the expected carbon emission reduction.

[0075] This calculation model transforms static carbon emission reduction calculations into dynamic, financially-oriented expected carbon emission reduction benefits by incorporating market elasticity and time value. This solves the problem of simply equating environmental benefits with fixed economic value, enabling scheduling decisions to respond more sensitively to carbon market dynamics and to more accurately integrate environmental value into economic objectives.

[0076] Subsequently, in step S3, the system enters the core decision-making stage. The predicted train braking energy, multi-vehicle energy conflict probability, health status prediction, and expected carbon emission reduction benefits obtained in step S2 are input into a decision model. This decision model considers the train (energy producer), energy storage system (energy storage / releaser), power grid (energy absorber / replenisher), and carbon trading node (environmental beneficiary) as four game participants with different interests. For example, the train operator wants braking energy to be efficiently recovered to avoid waste; the energy storage system operator wants to optimize charging and discharging to extend its lifespan and obtain charging and discharging price differences; the power grid operator wants to smooth load fluctuations; and the carbon trading node represents the social benefits of carbon emission reduction.

[0077] The decision-making model employs a multi-agent reinforcement learning algorithm that integrates game theory strategies (such as those based on...). The framework employs a multi-agent deep deterministic policy gradient algorithm, incorporating concepts of Nash equilibrium or cooperative game theory. Each participant is represented by an agent whose policy network outputs action suggestions (such as suggested charging / discharging power) based on the current system state (including the aforementioned input data). Through interaction with the environment (i.e., training on simulated or historical data), the agents not only learn to maximize their own long-term gains but also learn to interact and coordinate with other agents' policies using a game theory framework. The model, through collaborative optimization, ultimately aims to find an equilibrium strategy that maximizes the comprehensive gains of all participants (defined as a weighted sum including energy recovery gains, battery life depreciation costs, grid interaction costs / gains, carbon emission reduction gains, etc.). After training convergence, the decision model can quickly output optimal charging / discharging scheduling instructions (including target power, timing, etc.) for each energy storage system based on real-time input data. This mechanism solves the problem that traditional single-objective optimization or fixed-rule scheduling cannot balance the conflicting interests of multiple parties in complex dynamic environments.

[0078] Finally, in step S4, the system executes instructions through a cloud-edge-device collaborative scheduling architecture. Under normal circumstances, the cloud platform centrally runs complex prediction and decision-making models and forwards the generated charging and discharging scheduling instructions to the corresponding trackside or onboard energy storage system execution terminals via edge nodes (deployed at stations or regional control centers). When the edge node detects a network communication interruption with the cloud platform, the architectural advantages of this embodiment become apparent: the edge node immediately switches to autonomous operation mode. Based on locally cached lightweight versions of prediction models (such as simplified LSTM models and parameter-fixed degenerate models) and decision logic (such as rule-based or simplified game matrix-based emergency strategies), it autonomously generates and issues emergency scheduling instructions using locally available data. This ensures that even in the event of a communication failure, the energy storage system can still perform basic and reasonable scheduling based on local information, greatly improving the reliability and resilience of the entire scheduling system.

[0079] Technical effects of Example 1:

[0080] Improved Prediction Accuracy and Foresight: By employing a deep learning short-term prediction model that integrates special functions (error function and Bessel function), the braking energy and conflict risk of multiple vehicles are accurately quantified; and by using an electrochemical-environmental degradation model coupled with multiple stress factors, battery life is dynamically assessed. Both provide high-precision, multi-dimensional, and forward-looking key state inputs for optimized scheduling.

[0081] Dynamic collaborative optimization of multiple interests: By constructing a decision-making model based on multi-agent reinforcement learning and game theory, the goals of multiple interests such as energy recovery, equipment lifespan, grid interaction and dynamic carbon revenue are incorporated into a unified optimization framework, realizing comprehensive equilibrium optimization of economic, lifespan, stability and environmental benefits in complex dynamic environments.

[0082] Dynamic unity of economic and environmental benefits: By introducing a calculation model that includes carbon price elasticity factors and time effectiveness factors, real-time carbon prices and market expectations are dynamically transformed into carbon emission reduction benefits, so that environmental value can be accurately quantified and integrated into economic scheduling targets in real time. This guides the system to automatically respond to carbon market signals while pursuing economic benefits, thereby achieving dynamic green operation.

[0083] The system's reliability and resilience have been significantly enhanced: relying on a cloud-edge-device collaborative architecture, a combination of centralized intelligence and edge autonomy has been achieved. When cloud communication is interrupted, edge nodes can autonomously execute emergency scheduling based on local caching models and logic, ensuring the continuous operation of core scheduling functions under abnormal conditions and improving the availability and robustness of the entire energy management system.

[0084] Example 2 is the second embodiment of the present invention. This embodiment provides a method for scheduling regenerative braking energy storage based on dynamic prediction. Its basic architecture and data flow are the same as in Example 1, but the decision model in step S3 is further optimized and specified. This embodiment focuses on the use of a hybrid decision model based on a multi-agent deep deterministic policy gradient algorithm and a game tree search algorithm, and its detailed working principle.

[0085] Working principle of Example 2:

[0086] In step S3, the decision-making process is executed by a hybrid decision-making model. The core design goal of this model is to not only seek long-term optimization of global benefits in a highly dynamic and complex environment with intertwined interests, but also to conduct in-depth strategy deduction at key decision points to generate more robust and coordinated charging and discharging scheduling instructions.

[0087] The hybrid decision-making model consists of two collaborative components: a central evaluation network and multiple actuator networks. These actuator networks correspond to different game participants: the train (energy producer), the energy storage system (energy storage / releaser), the power grid (energy absorber / supplementer), and carbon trading nodes (environmental beneficiaries). The central evaluation network acts as both the "overall referee" and the "value guide."

[0088] The specific working principle is as follows:

[0089] Network Input and Role Localization: Actuator Network (Each Participant): Each actuator network makes decisions based on its corresponding local observation information. For example, the actuator network input for a train participant includes the train's real-time speed, acceleration, position, and received predicted train braking energy values. The components related to this vehicle; the actuator network inputs of the energy storage system participants include their own energy storage status data, health status predictions, and load information of associated lines.

[0090] Central Evaluation Network: Its input is global state information, integrating local observation information from all participants and all prediction and evaluation results output from step S2 (including predicted train braking energy values). Probability of energy conflict among multiple vehicles , Expected carbon emission reduction benefits ), and global power grid load data.

[0091] Reward function definition and strategy optimization loop: The behavior of each participant is driven by its reward function, which quantifies the immediate reward of each decision.

[0092] The revenue function for train stakeholders integrates instantaneous energy recovery efficiency with real-time operational safety indicators. A specific example is: (Energy recovery / Predicted braking energy) (Speed, tracking interval). The first item encourages efficient recovery of braking energy; the second item... It is a safety evaluation function. When the train's operating status (such as speed and distance from the train in front) approaches the safety boundary, the benefits will decrease. This prompts the train to prioritize operational safety in decision-making (such as by slightly adjusting the braking curve) rather than simply pursuing energy recovery. and These are the weighting coefficients.

[0093] The revenue function for energy storage system participants incorporates health status predictions. Determined lifetime attrition rate and including expected carbon reduction benefits The economic benefits. A specific example is: (Discharge revenue - Charging cost) The economic benefits are directly included in the carbon emission reduction benefits. (k is the conversion factor), which reflects the internalization of environmental value. It is a lifespan depletion cost function, which is related to the predicted health status. negative correlation (i.e.) The lower the value, the higher the "loss cost" per unit charge / discharge, and this is related to the charge / discharge current. ,temperature Positive correlation. and represents the weighting coefficient. This function enables the energy storage system to spontaneously adjust its charging and discharging strategies based on its own health status, proactively extending its lifespan while pursuing economic benefits.

[0094] The revenue functions for the power grid and carbon trading node participants are similarly defined, focusing on load smoothness and total carbon emission reduction, respectively.

[0095] The central evaluation network calculates a comprehensive reward value based on the global state and the actions of each participant. This value is not a simple sum of individual rewards, but rather reflects the overall optimization objective. Each actuator network then uses this comprehensive reward value to update its policy parameters through a multi-agent deep deterministic policy gradient algorithm, aiming to maximize its own long-term accumulated expected comprehensive reward value (judged by the central evaluation network). This process is a continuous learning and optimization loop, enabling each participant to approach Pareto optimality or Nash equilibrium through competition and cooperation.

[0096] Dynamic weight adjustment mechanism: To make the decision-making more adaptable to real-time conditions, this embodiment introduces a key mechanism: dynamically adjusting the weights of each game participant when the central evaluation network calculates the comprehensive reward value. Its triggering conditions and adjustment logic are set according to claim 7:

[0097] When the grid load exceeds a first load threshold (e.g., 85% of the regional grid peak load), the central evaluation network will increase the weight of "grid" participants in the comprehensive reward calculation. This means that dispatch decisions will be more inclined to prioritize absorbing braking energy to smooth peak and valley loads, or to reduce power extraction from the grid, thereby ensuring grid safety and stability.

[0098] When the predicted health status of an energy storage unit falls below a first health threshold (e.g., set to 0.8), the weight of that energy storage system participant in the overall reward will be reduced. This is equivalent to reducing the "voice" of the aging unit in the global optimization, guiding the system to reduce the charging and discharging stress allocation to it, thereby protecting vulnerable equipment and extending the overall system life.

[0099] When the probability of energy conflict between multiple trains exceeds a first conflict threshold (e.g., set to 0.7), the weight of conflict mitigation objectives among the participants in both the "train" and the "energy storage system" will be increased. This will drive the decision model to generate instructions for peak-shifting or pre-adjusting energy storage capacity, thereby preventing voltage fluctuations or energy waste caused by energy congestion.

[0100] This dynamic weight adjustment mechanism enables the hybrid decision-making model to adapt to different operating conditions, allowing it to flexibly adjust the focus of optimization under varying degrees of urgency or risk.

[0101] Game Tree Search and Final Instruction Generation: After the actuator network provides initial action strategies based on reinforcement learning, the hybrid decision model does not output directly. At critical decision moments (e.g., when the probability of multi-vehicle energy conflict exceeds a threshold, or when a large predicted value of train braking energy is expected), the model initiates a game tree search algorithm. This algorithm uses the current state as the root node and takes the action strategies suggested by each participant's actuator network as initial branches to perform multi-step deep deduction. At each future state node in the deduction, a lightweight payoff function and state transition model are invoked for evaluation. Through the search, the model can proactively analyze the possible game equilibrium results of different strategy combinations, identify and avoid strategy paths that, while offering high short-term gains, are detrimental in the long run or globally (e.g., instantaneous overload caused by all energy storage systems charging at full power simultaneously). Finally, the model selects the action strategy combination that achieves a more stable and coordinated game equilibrium and transforms it into specific charging and discharging scheduling instructions (e.g., instructing energy storage system A to charge at X kW for Y seconds; suggesting train B to brake slightly earlier, etc.).

[0102] Technical effects of Example 2:

[0103] Achieving better multi-objective collaboration and game equilibrium: Through the architecture of a central evaluation network and a multi-executor network, combined with a finely defined payoff function that integrates multiple dimensions such as safety, lifespan, economy, and environment, the decision-making model in this embodiment can more effectively coordinate the conflicting objectives of all parties. The scheduling strategy found is not only technically feasible, but also closer to the global optimal equilibrium point in terms of economic and social benefits.

[0104] Enhancing the robustness and foresight of decision-making: A game tree search algorithm is introduced as a post-optimization module for reinforcement learning strategies, enabling the decision-making process to possess deeper policy inference capabilities. This effectively avoids short-sighted decision-making or local optima problems that may exist in reinforcement learning models. Especially when dealing with high conflict probabilities (high probability of energy conflict among multiple vehicles) or large energy events (large predicted values ​​of train braking energy), it can predict and avoid systemic risks in advance, generating more robust scheduling instructions.

[0105] Enhancing the system's adaptive response to critical operating conditions: Through a dynamic weight adjustment mechanism based on explicit thresholds (such as the first load threshold, the first health threshold, and the first conflict threshold), the decision-making model is no longer static. It can keenly perceive grid pressure, equipment health risks, and energy conflict risks, and dynamically adjust the optimization focus, achieving a leap from "fixed strategy" to "contextual intelligence," significantly improving the system's operational safety and economy under abnormal or boundary conditions.

[0106] Balancing computational efficiency and decision quality: The hybrid model combines the advantages of deep reinforcement learning (excelling at handling continuous high-dimensional state-action spaces and performing long-term optimization) and game tree search (excelling at deep and accurate analysis of discrete strategies at critical nodes). Under normal operating conditions, it relies on an efficient learning network for rapid decision-making; under complex and critical operating conditions, it initiates a search for deep analysis. This hierarchical decision-making structure ensures both overall decision-making efficiency and decision quality at critical moments.

[0107] Example 3, the third embodiment of the present invention, provides a method for scheduling regenerative braking energy storage based on dynamic prediction. Its basic method flow, prediction model, and core architecture of the decision model are the same as in Examples 1 and 2. The optimization focus of this embodiment lies in a detailed explanation of the specific communication implementation method of the cloud-edge-device collaborative scheduling architecture in step S4, the accurate judgment criteria for communication interruption, and the emergency operation mechanism for edge nodes.

[0108] Working principle of Example 3:

[0109] In step S4, the system executes the charge / discharge scheduling instructions generated by the decision model. Its reliability highly depends on the stability and robustness of data transmission within the cloud-edge-device architecture. This embodiment features enhanced design for the communication link and edge autonomous logic.

[0110] Construction and Operating Mode of Dual-Mode Redundant Communication Link: The cloud-edge-device architecture employs a dual-mode redundant communication link consisting of 5G communication and BeiDou short message communication for the transmission of all critical data (including uploaded sensor data and issued scheduling commands). In normal operating mode, the system prioritizes the 5G communication link by default. 5G communication features high bandwidth and low latency, enabling it to efficiently carry large volumes of sensor data streams and complex scheduling commands, meeting the needs of centralized computing and real-time control in the cloud.

[0111] The BeiDou short message communication link serves as a hot backup link and is always in standby mode. Its core advantages lie in its independence from ground base stations, its ability to provide full coverage, and its strong resilience. The dual-mode link is managed through an intelligent routing module, ensuring that data packets are transmitted via the optimal or backup path.

[0112] Intelligent detection and handover of communication interruptions: Edge nodes (deployed at stations or regional control centers) continuously monitor the communication quality of their primary (5G) link with the cloud platform. Key monitored indicators include communication latency and signal strength. The system presets two key thresholds as benchmarks for determining communication anomalies:

[0113] First latency threshold: For example, set to 200 milliseconds. When the monitored average round-trip time consistently exceeds this threshold, it indicates network congestion or link instability, which may lead to severe delays in scheduling instructions.

[0114] First strength threshold: For example, set to -90dBm. When the received signal strength remains below this threshold, it indicates poor wireless channel conditions and a high risk of bit error rate or outage.

[0115] When an edge node detects that the communication latency with the cloud platform exceeds the first latency threshold and the signal strength is below the first strength threshold, the intelligent routing module determines that the primary link is unreliable and classifies it as a communication interruption. This AND logic avoids erroneous switching due to instantaneous fluctuations and improves the accuracy of status judgment. Once an interruption is determined, the system immediately and automatically switches the communication link to the BeiDou short message communication link to maintain the minimum exchange of critical information (such as simplified status reports and emergency command confirmations). More importantly, this judgment result directly triggers the edge node to start autonomous operation mode, beginning to generate and issue emergency dispatch commands locally.

[0116] Lightweight model and emergency logic for edge nodes: In order to achieve real-time computing on resource-constrained edge nodes, the prediction model cached therein is not the original complex model in the cloud, but a lightweight version obtained through model pruning (removing redundant connections and neurons in the neural network) and quantization (converting model parameters from high-precision floating-point numbers to low-precision fixed-point numbers).

[0117] A lightweight version of the short-term prediction model retains the core structure of the LSTM network, but significantly reduces the number of layers and neurons, and reduces the input features to the most critical ones (such as train speed and gradient in the local area). It can still output approximate braking energy trends and conflict risk indicators, but the computation speed is significantly improved.

[0118] A lightweight version of the lifespan prediction model: It may be simplified to a lookup table or linear decay model based on the current health status prediction, temperature, and current, to quickly estimate the aging cost for the next period.

[0119] The logic for generating emergency dispatch instructions is designed to prioritize ensuring the safety of train operation and energy storage equipment. Specific principles include:

[0120] Primary objective one: Ensure energy recovery during emergency braking of trains. When an emergency or routine braking state is detected or predicted, the edge node will instruct the associated energy storage system to unconditionally enter the "maximum safe power reception" mode, prioritizing the absorption of braking energy to prevent heat waste or brake shoe wear caused by regeneration failure, thus ensuring train braking safety.

[0121] The second primary objective is to ensure the safe operation of the energy storage system. When generating instructions, the system strictly adheres to locally cached safety boundary parameters for the energy storage system (such as maximum permissible charge / discharge power, voltage, SOC safety window, and stress limits corresponding to the current health state prediction value estimated based on the lightweight lifespan model). For example, for battery cabinets with a health state prediction value below 0.75, the emergency logic will significantly limit their charge / discharge current to avoid overstress impacts.

[0122] Simplified optimization objectives: Under the premise of meeting the above security objectives, edge nodes can perform limited peak shaving and valley filling optimization based on simplified local load information and electricity price information, but its complexity and economic pursuit are far lower than the global decision-making model in the cloud.

[0123] Technical effects of Example 3:

[0124] Revolutionary improvements in communication reliability and system availability: By introducing a dual-mode redundant communication link consisting of 5G and BeiDou short message service, the scheduling system in this embodiment possesses "integrated air-ground-space" communication capabilities. The 5G link ensures the requirements for large data volumes and low latency under normal circumstances; when extreme situations (such as natural disasters or infrastructure failures) cause the ground network to be paralyzed, the BeiDou link provides a crucial backup communication means, ensuring that the scheduling system remains connected even under the worst conditions, greatly improving the survivability and availability of the entire rail transit energy management system.

[0125] Precise communication status assessment and automated switching: By setting specific first latency thresholds (e.g., 200ms) and first intensity thresholds (e.g., -90dBm), and using AND logic for comprehensive judgment, the system can accurately distinguish between temporary network jitter and substantial interruptions, avoiding unnecessary frequent switching and malfunctions. Upon judgment, an emergency mode is automatically triggered, achieving a seamless and smooth transition from centralized cloud control to autonomous edge control, ensuring the continuity of scheduling services.

[0126] Edge intelligence achieves a balance between efficiency and security: By employing model pruning and quantization techniques, lightweight versions of predictive models are deployed at edge nodes, enabling critical state predictions even with limited computing resources, providing a basis for emergency decision-making. The emergency logic explicitly prioritizes safety, focusing on energy recovery and equipment protection. This allows the system to make the most conservative and reliable decisions in abnormal operating conditions such as communication interruptions, effectively preventing safety incidents that might arise from risky optimizations due to incomplete information.

[0127] A hierarchical and categorized elastic scheduling system was constructed: This embodiment, together with Embodiments 1 and 2, constitutes a complete elastic scheduling system. Under normal circumstances, the cloud performs global optimization scheduling based on full data and complex models (Embodiments 1 and 2); in the event of communication anomalies, edge nodes perform local backup scheduling based on lightweight models and simplified security rules (Embodiment 3). This hierarchical model of "global optimization + local backup" enables the system to maximize economic benefits while ensuring absolute basic security, possessing strong environmental adaptability and resilience.

[0128] A regenerative braking energy storage scheduling system based on dynamic prediction is proposed, applied to the above-mentioned regenerative braking energy storage scheduling method based on dynamic prediction, with reference to... Figure 2 ,include:

[0129] Data acquisition module 1 is used to collect vehicle operation data, line condition data, environmental status data, energy storage status data, power grid load data, and real-time carbon price data;

[0130] Collaborative prediction module 2, connected to data acquisition module 1, is used to input vehicle operation data, line condition data, and power grid load data into a short-term prediction model to generate train braking energy prediction values ​​and multi-vehicle energy conflict probabilities for a future preset period; input energy storage status data and environmental status data into a lifetime prediction model to generate energy storage system health status prediction values; and calculate the expected carbon emission reduction benefits based on train braking energy prediction values ​​and real-time carbon price data.

[0131] Hybrid decision module 3, connected to collaborative prediction module 2, is used to input the predicted value of train braking energy, the probability of multi-vehicle energy conflict, the predicted value of health status, and the expected carbon emission reduction benefits into a decision model. The decision model takes the train, energy storage system, power grid and carbon trading node as game participants, integrates game theory strategies through multi-agent reinforcement learning algorithm, collaboratively optimizes the comprehensive benefits of each participant, and outputs charging and discharging scheduling instructions for the energy storage system.

[0132] The instruction execution and emergency module 4 is connected to the hybrid decision-making module 3 and the data acquisition module 1, respectively. Through the cloud-edge-device architecture of the collaborative scheduling system, it sends the charging and discharging scheduling instructions to the corresponding energy storage system for execution. When the edge node detects a communication interruption with the cloud platform, it autonomously generates and sends out emergency scheduling instructions based on the locally cached prediction model and decision logic.

[0133] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A regenerative braking energy storage scheduling method based on dynamic prediction, applied to a collaborative scheduling system including a cloud platform, edge nodes, and on-board and trackside execution terminals, characterized in that, The method includes: Step S1: Collect vehicle operation data through sensors deployed on the train, collect track condition data and environmental status data through sensors deployed along the track, collect energy storage status data through the battery management system of the energy storage cabinet, and obtain grid load data and real-time carbon price data through the data interface. Step S2: Input the vehicle operation data, the line condition data, and the power grid load data into a short-term prediction model to generate a predicted value for train braking energy and the probability of multi-vehicle energy conflict for a future preset period; input the energy storage status data and the environmental status data into a lifetime prediction model to generate a predicted value for the health status of the energy storage system; and calculate the expected carbon emission reduction benefits based on the predicted value for train braking energy and the real-time carbon price data. In step S2, the short-term prediction model is a deep learning model built on a long short-term memory network. Its input feature dimension is N, including the real-time speed and acceleration in the vehicle operation data, the slope and curvature in the road condition data, and the historical braking energy sequence, where N is an integer greater than 1. The deep learning model is used to output the predicted value of the train braking energy and the probability of multi-vehicle energy conflict within a first preset time period in a rolling prediction manner; wherein, the predicted value of the train braking energy Calculated from a regression subnetwork in the deep learning model: ； in, Let be the error function. It is a zero-order Bessel function. It is the natural logarithm function. This is a calibration coefficient for the system's energy conversion efficiency, used to calibrate the overall energy conversion efficiency of the formula. The number of trains participating in the prediction. For the first The train's braking force request characteristic coefficients are used to map real-time acceleration to weights for braking force requests. For the first Real-time acceleration data collected by the train's acceleration sensors. For the first The train speed fluctuation weighting factor is used to control the scaling of the difference between the real-time speed and the reference speed onto the Bessel function input. For the first The difference between the train's real-time speed and the reference speed. For the current moment, The duration of the first preset duration, This is a time-varying power density function reconstructed based on real-time speed and slope data, used to characterize the power density at time t. The instantaneous power change rate, The integral variable in the formula is denoted as ; the range of the predicted train braking energy is [0, +∞), and the unit is kilowatt-hour (kWh); The probability of multi-vehicle energy conflict is calculated by a classification subnetwork in the deep learning model based on the same input features; Step S3: Input the predicted value of the train braking energy, the probability of energy conflict among multiple vehicles, the predicted value of the health status, and the expected carbon emission reduction benefits into a decision model; The decision model takes the train, energy storage system, power grid, and carbon trading node as game participants, integrates game theory strategies through a multi-agent reinforcement learning algorithm, coordinates and optimizes the comprehensive benefits of each participant, and outputs charging and discharging scheduling instructions for the energy storage system. Step S4: Through the cloud-edge-device architecture of the collaborative scheduling system, the charging and discharging scheduling command is sent to the corresponding energy storage system for execution; wherein, when the edge node detects a communication interruption with the cloud platform, it autonomously generates and sends out an emergency scheduling command based on the locally cached prediction model and decision logic.

2. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 1, characterized in that: In step S2, the lifetime prediction model is an electrochemical-environmental coupled degradation model. This model is used to dynamically correct the lithium-ion diffusion coefficient and interface impedance growth rate within the battery based on the battery's state of charge, health status, and temperature from the energy storage state data, as well as the humidity and electromagnetic intensity from the environmental state data. Based on the corrected parameters, it predicts the health status of the energy storage system. : ； in, It is an exponential function. This represents the nominal value of the initial health state of the energy storage system. The cumulative operating time of the energy storage system. The temperature data is from the environmental condition data. The humidity data is from the environmental status data. The temperature, humidity, and stress coupling function is defined as follows: ,in and The material aging response coefficient. For reference to ambient temperature, For reference to ambient temperature and humidity, It is the natural logarithm function. The charging and discharging current data in the energy storage state data. The electromagnetic intensity data is from the environmental state data. The electrochemical-electromagnetic stress function is defined as follows: ,in The electrochemical reaction rate constant is Electromagnetic interference factor, Let be the error function. This is the reference value for electromagnetic intensity. The integral variable in this formula; the predicted health status value The range of is (0,1], and it is a dimensionless ratio.

3. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 1 or 2, characterized in that: In step S2, the expected carbon emission reduction benefit is calculated as follows: the predicted value of train braking energy is converted into expected recyclable electrical energy, and combined with the real-time carbon price data and the real-time carbon quota information obtained from the carbon trading market, a calculation model that includes a carbon price elasticity factor and time validity is used for dynamic calculation.

4. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 1, characterized in that: In step S3, the decision model is a hybrid decision model constructed based on the multi-agent deep deterministic policy gradient algorithm and the game tree search algorithm. The hybrid decision-making model includes a central evaluation network and multiple executor networks corresponding to different game participants. The central evaluation network is used to output a comprehensive reward value based on global state information and the payoff function of each game participant. Each executor network is used to optimize the strategy based on local observation information and the comprehensive reward value and output the action strategy. Finally, the charging and discharging scheduling instruction is generated based on the game equilibrium result of each action strategy.

5. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 4, characterized in that: The payoff function of the train as a game participant integrates instantaneous energy recovery efficiency and real-time operational safety indicators, while the payoff function of the energy storage system as a game participant integrates the lifespan attrition rate determined based on the predicted health status value and the economic benefits including the expected carbon emission reduction benefits.

6. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 4, characterized in that: In step S3, when the decision model is running, it dynamically adjusts the weights of each game participant in the calculation of the comprehensive reward value based on at least one of the following conditions: whether the grid load exceeds the first load threshold, whether the predicted health status value is lower than the first health threshold, or whether the probability of multi-vehicle energy conflict is higher than the first conflict threshold.

7. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 5, characterized in that: In step S4, the cloud-edge-device architecture uses a dual-mode redundant communication link composed of 5G communication and Beidou short message communication for data transmission. When the edge node detects that the communication delay with the cloud platform exceeds the first delay threshold and the signal strength is lower than the first strength threshold, it determines that the communication is interrupted and triggers the local generation of the emergency dispatch command.

8. The regenerative braking energy storage scheduling method based on dynamic prediction according to claim 7, characterized in that: The prediction model cached locally at the edge node is a lightweight version of the short-term prediction model and the lifetime prediction model obtained through model pruning and quantization techniques; the generation logic of the emergency dispatch command prioritizes ensuring the safe operation of the train's emergency braking energy recovery and energy storage system.

9. A regenerative braking energy storage scheduling system based on dynamic prediction, applied to the regenerative braking energy storage scheduling method based on dynamic prediction as described in any one of claims 1-8, characterized in that, include: The data acquisition module (1) is used to collect vehicle operation data, line condition data, environmental status data, energy storage status data, power grid load data and real-time carbon price data; The collaborative prediction module (2) is connected to the data acquisition module (1) and is used to input the vehicle operation data, the line condition data and the power grid load data into a short-term prediction model to generate the train braking energy prediction value and the probability of multi-vehicle energy conflict for a future preset period. The energy storage status data and the environmental status data are input into a lifetime prediction model to generate a health status prediction value for the energy storage system; based on the predicted train braking energy value and the real-time carbon price data, the expected carbon emission reduction benefit is calculated. The hybrid decision module (3) is connected to the collaborative prediction module (2) and is used to input the predicted value of the train braking energy, the probability of energy conflict among multiple vehicles, the predicted value of the health status and the expected carbon emission reduction benefits into a decision model. The decision model takes the train, energy storage system, power grid and carbon trading node as game participants, integrates game theory strategies through multi-agent reinforcement learning algorithm, collaboratively optimizes the comprehensive benefits of each participant, and outputs charging and discharging scheduling instructions for the energy storage system. The instruction execution and emergency module (4) is connected to the hybrid decision module (3) and the data acquisition module (1) respectively. Through the cloud-edge-end architecture of the collaborative scheduling system, the charging and discharging scheduling instruction is sent to the corresponding energy storage system for execution. When the edge node detects a communication interruption with the cloud platform, it autonomously generates and sends an emergency scheduling instruction based on the local cached prediction model and decision logic.

Images