Mobile energy storage vehicle energy management method based on intelligent algorithm

Abstract

The invention relates to the technical field of mobile energy storage vehicle energy management, and discloses a mobile energy storage vehicle energy management method based on an intelligent algorithm, and the method comprises the steps: collecting the charging and discharging rate of a battery pack, the environment temperature, the power grid load fluctuation and other multi-dimensional energy state data; a heterogeneous layered architecture of edge computing nodes, a cloud collaboration platform and a vehicle-mounted control terminal is constructed, energy partitions are divided according to peak valleys of a power grid, and low-delay response, global optimization and local closed-loop control modes are configured; fusing the multi-source heterogeneous energy data streams and removing abnormal values; a dynamic optimization decision model is constructed through a battery degradation model and a power grid supply and demand balance equation, and a multi-parameter collaborative constraint relation is solved through simultaneous solving of a multi-target iteration solver and a parallel gradient descent algorithm; and generating a three-level adaptive response instruction sequence of local battery overload alarm, regional power grid frequency modulation early warning and global energy scheduling imbalance pre-judgment based on constraint boundary triggering conditions. According to the method, the accuracy, the collaboration and the robustness of energy management are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy management of mobile energy storage vehicles, and in particular to an energy management method for mobile energy storage vehicles based on an intelligent algorithm. Background Art

[0002] With the rapid development of the new energy industry, mobile energy storage vehicles, as flexible energy storage and regulation units, are increasingly used in scenarios such as power grid peak shaving and emergency power supply. However, the current energy management of mobile energy storage vehicles faces many challenges, and more efficient management methods are urgently needed to improve their operating performance and grid coordination capabilities.

[0003] From the perspective of energy data management, traditional methods can only process energy data of a single dimension, and it is difficult to comprehensively consider the coupling effects of multiple factors such as battery pack charge and discharge rate, ambient temperature distribution, and grid load fluctuations. For example, ignoring the significant effect of ambient temperature on battery performance may lead to a decrease in battery charge and discharge efficiency or even increased life loss at extreme temperatures; and without fully analyzing the characteristics of grid load fluctuations, it is impossible to accurately grasp the energy scheduling needs during peak and valley periods, making it difficult to achieve efficient interaction between mobile energy storage vehicles and the grid.

[0004] In terms of system architecture deployment, existing energy management systems mostly adopt a centralized control architecture, which has problems such as high response delay and insufficient global optimization capabilities. Under the centralized architecture, the synergy between edge devices and cloud platforms is poor, and it is impossible to flexibly perform energy partition scheduling according to the peak and valley periods of grid load fluctuations. For example, during the peak load period of the grid, the lack of an effective hierarchical scheduling mechanism may cause mobile energy storage vehicles to be unable to respond to the grid frequency modulation needs in a timely manner, affecting the stability of the grid.

[0005] In the data processing and optimization decision-making link, traditional algorithms are difficult to cope with the fusion processing and dynamic optimization solution of multi-source heterogeneous data. Multi-source data (such as current time series data, load frequency data, temperature sampling data) have problems such as sampling interval deviation and noise interference. Traditional data cleaning and fusion methods are difficult to ensure temporal and spatial consistency, resulting in insufficient input data accuracy for the optimization decision model. At the same time, most existing optimization models are single-objective optimization (such as only considering energy efficiency or thermal management), and cannot jointly solve the collaborative constraint relationship of multiple variables such as charging and discharging efficiency, temperature compensation, and load distribution, making it difficult to achieve the global optimality of energy management.

[0006] In terms of response mechanism, the traditional threshold-triggered alarm method lacks adaptive adjustment capability and cannot adjust the response strategy in time according to the dynamic changes in the system operation status. For example, a single current threshold alarm cannot distinguish between instantaneous current fluctuations and real overload conditions, which may lead to false alarms or missed alarms; for power grid frequency regulation warnings and global energy scheduling imbalance prediction, there is a lack of multi-scale analysis and early prediction capabilities, making it difficult to take effective control measures before the system becomes unstable.

[0007] In addition, with the large-scale application of mobile energy storage vehicles, the complexity of their interaction with the power grid has increased significantly, and traditional energy management methods are insufficient in terms of scalability and robustness. How to build a scalable heterogeneous layered architecture to achieve the collaborative work of edge computing nodes, cloud platforms and vehicle terminals, and how to design efficient optimization algorithms to adapt to large-scale data processing and real-time decision-making needs have become technical problems that need to be solved urgently.

[0008] The existing energy management methods for mobile energy storage vehicles have obvious defects in multi-dimensional data processing, layered architecture coordination, dynamic optimization decision-making and adaptive response mechanism, which makes it difficult to meet the smart grid's demand for efficient, reliable and flexible energy management of mobile energy storage systems. Therefore, it is urgent to propose an energy management method for mobile energy storage vehicles based on intelligent algorithms, which can improve the energy utilization efficiency, grid coordination ability and system operation reliability of mobile energy storage vehicles through multi-technology integration and architecture innovation. Summary of the invention

[0009] The purpose of the present invention is to provide an energy management method for a mobile energy storage vehicle based on an intelligent algorithm to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: an energy management method for a mobile energy storage vehicle based on an intelligent algorithm, the method comprising:

[0011] Collect multi-dimensional energy status data of the mobile energy storage vehicle, wherein the multi-dimensional energy status data includes the real-time charge and discharge rate of the battery pack, the ambient temperature distribution characteristics, and the grid load fluctuation characteristics;

[0012] Based on the dynamic response requirements of the energy scheduling unit, a heterogeneous layered deployment architecture of edge computing nodes, cloud collaboration platform and vehicle control terminal is generated;

[0013] Synchronously acquiring the real-time energy interaction data stream of each node under the heterogeneous layered deployment architecture, and performing multi-source heterogeneous data fusion and outlier elimination processing on the real-time energy interaction data stream;

[0014] A dynamic optimization decision model is constructed by using a battery degradation model and a power grid supply and demand balance equation, and a coordinated constraint relationship among a charge and discharge efficiency matrix, a temperature compensation coefficient tensor, and a load distribution vector in the dynamic optimization decision model is solved jointly;

[0015] Based on the boundary triggering conditions of the collaborative constraint relationship, a three-level adaptive response instruction sequence is generated, including local battery overload alarm, regional power grid frequency regulation warning and global energy scheduling imbalance prediction.

[0016] Preferably, the heterogeneous hierarchical deployment architecture is characterized by a distributed computing layer; the heterogeneous hierarchical deployment architecture of the generated edge computing nodes, cloud collaborative platform and vehicle-mounted control terminal includes:

[0017] According to the peak-valley period division strategy in the grid load fluctuation characteristics, define the energy partition rules for the core scheduling area, buffer scheduling area and standby scheduling area;

[0018] Based on the energy partition rules, configure a low-latency response mode for the edge computing nodes, a global optimization mode for the cloud collaborative platform, and a local closed-loop control mode for the vehicle-mounted control terminal.

[0019] Preferably, the dynamic optimization decision model is implemented by a multi-objective iterative solver; the simultaneous solution of the collaborative constraint relationships of the charge-discharge efficiency matrix, temperature compensation coefficient tensor and load distribution vector in the dynamic optimization decision model includes:

[0020] Load the charge-discharge efficiency matrix into the energy conservation equations of the multi-objective iterative solver, and update the capacity decay factor based on the battery health state model;

[0021] Embed the temperature compensation coefficient tensor into the thermodynamic diffusion equation, and calculate the anisotropic thermal conductivity through Fourier's law;

[0022] Discretize the load distribution vector into a time series task queue, and construct a task constraint matrix in combination with the dynamic priority scheduling strategy;

[0023] Use the parallel gradient descent algorithm to synchronously update the optimization residuals of the energy conservation equations, thermodynamic diffusion equation and task constraint matrix until the steady-state convergence condition is reached.

[0024] Preferably, the multi-objective iterative solver includes an energy efficiency optimization module, a thermal management module and a task scheduling module; the energy efficiency optimization module, thermal management module and task scheduling module are cross-module coupled through a state sharing interface;

[0025] A battery life prediction model is configured in the energy efficiency optimization module, a phase change material thermal buffer model is configured in the thermal management module, and a real-time preemptive scheduling algorithm is configured in the task scheduling module;

[0026] The state sharing interface transfers the charge-discharge efficiency correction amount, temperature gradient correction amount and task priority correction amount to the adjacent functional modules during each iteration.

[0027] Preferably, the parallel gradient descent algorithm is implemented by a phased learning rate adjustment strategy; the use of the parallel gradient descent algorithm to synchronously update the optimization residuals of the energy conservation equations, thermodynamic diffusion equation and task constraint matrix includes:

[0028] Take the charge-discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector of the current iteration step as the initial optimization input;

[0029] Calculate the energy residual norm through the energy efficiency optimization module, calculate the thermal balance residual norm through the thermal management module, and calculate the scheduling delay residual norm through the task scheduling module;

[0030] According to the dynamic weighted geometric mean of the three types of residual norms, adaptively adjust the learning rate decay coefficient of the next iteration step;

[0031] When the dynamic weighted geometric mean is lower than the preset optimization threshold, terminate the iteration and output the optimized decision parameters.

[0032] Preferably, the adjustment strategy of the learning rate decay coefficient is implemented based on the adaptive momentum mechanism; the adaptive adjustment of the learning rate decay coefficient of the next iteration step includes:

[0033] If the convergence rate of the residual norm in the current iteration step is higher than the rate threshold, multiply the learning rate amplification factor by the reference learning rate as the new parameter;

[0034] If the convergence rate of the residual norm is in a flat interval, multiply the learning rate decay factor by the reference learning rate as the new parameter;

[0035] If the residual norm shows periodic fluctuations, enable the random restart strategy to reset the optimization direction and step size range.

[0036] Preferably, the three-level adaptive response instruction sequence is generated by an event-driven state machine; the generation of the three-level adaptive response instruction sequence including local battery overload warning, regional power grid frequency modulation warning, and global energy scheduling imbalance prediction includes:

[0037] Set a primary trigger condition based on current mutation for local battery overload warning, and use the sliding time window statistical method to detect the instantaneous current abnormal peak;

[0038] Set a secondary trigger condition based on frequency deviation for regional power grid frequency modulation warning, and identify the power grid harmonic distortion area through the multi-scale frequency domain analysis algorithm;

[0039] Set a final trigger condition based on load deviation exceeding the limit for global energy scheduling imbalance prediction, and use the energy distribution entropy value calculation to determine the instability critical point of the system.

[0040] Preferably, the event-driven state machine includes nested conditional jump logic; the use of the sliding time window statistical method to detect the instantaneous current abnormal peak includes:

[0041] Extract the mean, range, and kurtosis features of the current monitoring data within a fixed time window;

[0042] Fitting the probability distribution function of the current peak value through the generalized Pareto distribution model;

[0043] When the observed current value exceeds the upper limit of the 99.9% confidence interval of the probability distribution function, trigger the local battery overload warning instruction.

[0044] Preferably, the multi-scale frequency domain analysis algorithm is implemented by using the wavelet transform decomposition technology; the identification of the power grid harmonic distortion area through the multi-scale frequency domain analysis algorithm includes:

[0045] Mapping the power grid frequency fluctuation signal into a multi-resolution wavelet coefficient matrix, and extracting the high-frequency abnormal components in the matrix;

[0046] Based on the clustering analysis algorithm, merging the high-amplitude coefficient clusters with spatio-temporal correlation to generate the potential harmonic interference area markers;

[0047] Calculating the product index of the energy proportion and frequency offset of each marked area, and screening the target intervals that meet the preset frequency modulation level.

[0048] Preferably, the multi-source heterogeneous data fusion and outlier rejection processing is implemented by using a robust data cleaning strategy; the multi-source heterogeneous data fusion and outlier rejection processing of the real-time energy interaction data stream includes:

[0049] Establishing the data mapping relationship among the current time series of the edge computing node, the load frequency series of the cloud platform, and the temperature sampling series of the vehicle-mounted terminal;

[0050] Separating the high-frequency noise components in the data stream of each node through the empirical mode decomposition algorithm, and performing cross-modal alignment on the low-frequency effective data based on the robust principal component analysis algorithm;

[0051] Using the time series interpolation algorithm to compensate for the sampling interval deviation between multi-source data streams, and generating a spatio-temporally consistent fusion energy management tensor.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] In terms of multi-dimensional energy data management, by collecting multi-dimensional data such as the real-time charge and discharge rate of the battery pack, the environmental temperature distribution characteristics, and the power grid load fluctuation characteristics, and adopting a robust data cleaning strategy (such as empirical mode decomposition, robust principal component analysis, and time series interpolation algorithm) for multi-source heterogeneous data fusion and outlier rejection, the limitations of the traditional single data processing mode are effectively solved. This method can not only remove high-frequency noise, compensate for the sampling interval deviation, but also generate a spatio-temporally consistent fusion energy management tensor, providing a high-precision data basis for subsequent optimization decisions and improving the accuracy of energy state assessment.

[0054] The innovative design of the heterogeneous hierarchical deployment architecture is an important breakthrough of the present invention. By defining the energy partitioning rules of the core dispatching area, buffer dispatching area and backup dispatching area, and configuring low-latency response mode, global optimization mode and local closed-loop control mode for edge computing nodes, cloud collaboration platform and on-board control terminal respectively, a heterogeneous hierarchical architecture at the distributed computing level is constructed. This architecture enables the system to dynamically adjust the dispatching strategy according to the peak and valley periods of the power grid load: the edge node can quickly respond to local energy demand (such as instantaneous overload alarm), the cloud platform performs global energy optimization scheduling (such as cross-regional load distribution), and the on-board terminal ensures the real-time performance of local control, which significantly improves the response speed and global optimization capability of the system, and enhances the flexibility of coordinated dispatching between mobile energy storage vehicles and power grids.

[0055] The construction and solution of the dynamic optimization decision model realizes multi-objective collaborative optimization. By combining the battery degradation model with the grid supply and demand balance equation, a dynamic optimization decision model including the charge and discharge efficiency matrix, the temperature compensation coefficient tensor and the load distribution vector is constructed, and solved using a multi-objective iterative solver (integrated energy efficiency optimization, thermal management and task scheduling modules) and a parallel gradient descent algorithm. This method can not only synchronously update the optimization residuals of energy conservation, thermodynamic diffusion and task constraints, but also adjust the learning rate through an adaptive momentum mechanism, thereby realizing multi-objective collaboration of energy efficiency improvement, thermal management optimization and load balancing distribution. For example, the battery life in the energy efficiency optimization module

[0056] The life prediction model can extend the battery cycle life, the phase change material thermal buffer model of the thermal management module can reduce the damage to the battery caused by temperature fluctuations, and the real-time preemptive algorithm of the task scheduling module can improve the load distribution efficiency, thereby improving the energy utilization efficiency and system operation reliability as a whole.

[0057] The generation mechanism of the three-level adaptive response instruction sequence significantly enhances the robustness and safety of the system. Through the event-driven state machine and nested conditional jump logic, primary, secondary and final trigger conditions are set for local battery overload, regional power grid frequency regulation and global energy scheduling imbalance, and sliding time window statistics, multi-scale frequency domain analysis (wavelet transform) and energy distribution entropy value calculation are used to achieve accurate detection. For example, current peak detection based on generalized Pareto distribution can accurately identify overload anomalies, wavelet transform combined with cluster analysis can locate harmonic interference areas, and energy distribution entropy can predict the critical point of system instability. This hierarchical response mechanism can not only trigger alarms in time, but also dynamically adjust the response strategy according to the degree of abnormality, effectively reduce false alarms and missed alarms, and improve the system's adaptive ability and fault warning level under complex working conditions.

[0058] In addition, multi-module coupling and cross-layer collaborative design (such as passing correction amounts through state sharing interfaces and adjusting the learning rate in stages for the parallel gradient descent algorithm) further improve the system integration and optimization efficiency. Through real-time data sharing and parameter collaborative update among functional modules, a closed-loop optimization feedback mechanism is formed, making the energy management process more intelligent and efficient, providing technical support for the large-scale application of mobile energy storage vehicles in smart grids, and having significant economic value and social significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the working principle diagram of the energy management method for mobile energy storage vehicles based on intelligent algorithms according to the present invention;

[0060] Figure 2 It is the solution flowchart of the dynamic optimization decision model;

[0061] Figure 3 It is the structural diagram of the multi-objective iterative solver;

[0062] Figure 4 It is the generation diagram of the three-level adaptive response instruction;

[0063] Figure 5 It is the processing diagram of multi-source heterogeneous data fusion. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] Please refer to Figures 1 - 5 , the energy management method for mobile energy storage vehicles based on intelligent algorithms involved in the present invention is specifically implemented as follows:

[0066] Through on-vehicle sensors and grid access devices, the real-time charge and discharge rate of the battery pack of the mobile energy storage vehicle, the environmental temperature distribution characteristics, and the grid load fluctuation characteristics are collected in real time. Among them, the charge and discharge rate of the battery pack is obtained through current sensors and power measurement modules, the environmental temperature distribution characteristics are collected through a temperature sensor network arranged at key positions such as the battery compartment and the motor compartment, and the grid load fluctuation characteristics are obtained by synchronizing the grid load curve and the peak-valley period division information through the grid data interface in real time.

[0067] Based on the dynamic response requirements of the energy scheduling unit, a heterogeneous hierarchical deployment architecture consisting of edge computing nodes, cloud collaborative platforms, and vehicle-mounted control terminals is constructed. This architecture realizes the functional division of different levels through a distributed computing hierarchy to meet the dual requirements of real-time performance and global optimization.

[0068] Synchronously obtain the real-time energy interaction data streams of each node under the heterogeneous hierarchical deployment architecture, including the current time series data of edge computing nodes, the load frequency data of cloud platforms, and the temperature sampling data of vehicle-mounted control terminals. Perform multi-source heterogeneous data fusion and outlier removal processing on the above data streams. By establishing cross-node data mapping relationships, separating high-frequency noise, cross-modal aligning low-frequency valid data, and compensating for sampling interval deviations, a spatio-temporally consistent fusion energy management tensor is generated.

[0069] Construct a dynamic optimization decision model through a battery degradation model and a power grid supply-demand balance equation, and jointly solve the collaborative constraint relationships of the charge-discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector in the model. Specifically, the model is solved through a multi-objective iterative solver, comprehensively considering energy conservation, thermodynamic diffusion, and task scheduling constraints to optimize the charge-discharge strategy and energy distribution of the energy storage vehicle.

[0070] Based on the boundary trigger conditions of the collaborative constraint relationship, generate a three-level adaptive response instruction sequence including local battery overload warnings, regional power grid frequency modulation warnings, and global energy scheduling imbalance predictions. The instruction generation is realized through an event-driven state machine, and different trigger conditions and detection methods are set for warnings and warnings at different levels.

[0071] Example 1:

[0072] In the process of implementing a heterogeneous hierarchical deployment architecture, it is first necessary to define the energy partition rules for the core scheduling area, buffer scheduling area, and standby scheduling area based on the peak-valley period division strategy in the power grid load fluctuation characteristics. The power grid load fluctuation characteristics are obtained through statistical analysis of historical load data. The peak-valley period division can be determined by the time-of-use electricity price mechanism or load forecasting algorithms. For example, through Fourier transform for frequency-domain analysis of the power grid load curve, the daily load peak period (such as the peak industrial electricity consumption during the day), valley period (such as the low residential load period at night), and flat period can be identified. The core scheduling area corresponds to the high-load area or key power supply nodes during the peak-valley period, and it is necessary to prioritize ensuring the stability of energy supply. For example, the central business district and industrial parks in the city. In this area, the real-time response demand for energy storage vehicles during the load peak is extremely high, and any interruption of energy supply may cause significant economic losses or service interruptions. The buffer scheduling area is located between the core scheduling area and the standby scheduling area, playing the role of energy transfer and regulation. Areas with medium load density and central geographical location can be selected, such as the outskirts of the city or around transportation hubs. Its function is to quickly supplement energy when the energy demand in the core scheduling area surges, or to temporarily store excess energy before the standby scheduling area is activated, avoiding drastic fluctuations in energy supply and demand. The standby scheduling area, as the redundant reserve of the system, is usually set in areas far from the core load center but with good transportation and energy access conditions, such as suburban energy storage stations or areas rich in renewable energy. It is only activated when both the core scheduling area and the buffer scheduling area are unable to meet the energy demand to cope with extreme load events or large-scale equipment failures.

[0073] Based on the above energy partition rules, different working modes are configured for the edge computing nodes, cloud collaborative platforms, and vehicle-mounted control terminals respectively. The edge computing nodes are deployed at physical locations close to the energy storage vehicle or the grid access point, such as the edge computing devices of in-vehicle edge servers or distributed energy sites. Their hardware architectures usually include high-performance processors, real-time operating systems, and high-speed communication modules. For the low-latency response requirements of the core scheduling area, a millisecond-level real-time operating system (such as QNX or VxWorks) is configured for the edge computing nodes, and their data processing flow is optimized: when receiving a sudden change signal in the charging and discharging rate of the battery pack (such as the current change rate exceeding a preset threshold), the edge computing node needs to complete data acquisition, anomaly detection, and control instruction generation within 10 milliseconds. The specific implementation method is as follows: periodic data sampling is triggered by a hardware timer, the sampling period is set to 1 millisecond, and digital filtering (such as finite impulse response filtering) is immediately performed after each sampling to remove high-frequency noise. Subsequently, an anomaly detection algorithm based on threshold comparison is executed. If an anomaly is detected, a predefined control strategy library (such as proportional-integral-derivative control algorithm) is immediately called to generate an instruction to adjust the charging and discharging power, and the instruction is sent to the vehicle-mounted actuator through a hard real-time communication interface (such as CAN bus or industrial Ethernet) to ensure that the instruction transmission delay is less than 5 milliseconds. For the edge computing nodes in the buffer scheduling area and the standby scheduling area, the response time can be appropriately relaxed to the 100-millisecond level to reduce hardware costs and computing resource consumption. Their data processing flow can add data caching and batch processing links, such as collecting data once every 100 milliseconds and performing sliding window averaging processing to reduce over-response to sudden noise.

[0074] As the core of global energy optimization, the cloud collaborative platform is deployed in the cloud computing data center and has the capabilities of large-scale data storage, distributed computing, and cross-regional coordination. Its software architecture adopts the microservices design pattern, including multiple independent service modules such as energy prediction, optimal scheduling, and security monitoring. To meet the global optimization requirements of the core scheduling area, the cloud collaborative platform needs to integrate the real-time data of multiple edge computing nodes (such as the status information of hundreds to thousands of mobile energy storage vehicles), predict the grid load trend and the energy state of energy storage vehicles in the next 15 minutes to 1 hour through machine learning models (such as long short-term memory networks), and solve the global energy optimization problem based on the mixed integer programming algorithm. The objective functions include minimizing the energy transmission cost and maximizing the smoothing of the peak-valley difference of the grid. For example, 30 minutes before the peak load arrives, the cloud collaborative platform sends centralized charging or discharging instructions to the mobile energy storage vehicles around the core scheduling area according to the prediction results. By optimizing the charging and discharging power distribution of each vehicle, the fluctuation range of the grid load is reduced by more than 20% (this is only a logical description here and does not involve specific experimental data). For the buffer scheduling area, the optimization period of the cloud collaborative platform can be extended to 1 hour, focusing on the energy balance between regions. For example, by coordinating the energy interaction between energy storage vehicles and distributed photovoltaic and wind farms in the buffer scheduling area, the energy self-sufficiency within the region can be achieved. The cloud collaborative function in the standby scheduling area is mainly used for long-term energy planning, such as adjusting the charging plan of energy storage vehicles in the standby scheduling area according to the seasonal energy demand changes to ensure sufficient energy reserves in extreme cases.

[0075] The on-vehicle control terminal is the core unit for realizing the local energy closed-loop control of the energy storage vehicle. It is usually integrated into the vehicle's central control system and includes a microcontroller, a sensor interface, an actuator drive circuit, and a human-machine interface. For the local closed-loop control mode, the on-vehicle control terminal needs to collect parameters such as the voltage, current, and temperature of the battery pack in real time, and adjust the actuators such as the charge and discharge switches and the cooling fan speed of the battery management system (BMS) through a PID controller or a model predictive control algorithm to keep the battery pack operating within a safe range. The specific control process is as follows: The temperature sensor network (such as DS18B20 digital temperature sensors) collects the temperature of each battery cell in real time and transmits the data to the on-vehicle control terminal every 50 milliseconds. The terminal first fuses the multi-sensor data through a Kalman filter algorithm to eliminate measurement noise, and then calculates the average temperature and temperature gradient of the battery pack. If the average temperature exceeds 25°C (the preset threshold), the cooling fan is started, and the speed of each fan is adjusted according to the temperature gradient to make the temperature field distribution of the battery pack uniform; if the temperature continues to rise to 35°C, a power reduction charging strategy is triggered, and the charge and discharge power is reduced by 30% to reduce heat generation. At the same time, the on-vehicle control terminal also needs to communicate with the edge computing node and the cloud collaborative platform, receive upper-layer scheduling instructions, and feedback local status information. For example, when receiving an emergency discharge instruction from the edge computing node, it needs to switch to the discharge mode within 50 milliseconds and report the discharge power and the remaining battery capacity in real time.

[0076] In the coordinated operation of the energy partition rule and the node working mode, data interaction is realized between each layer through a standardized communication protocol (such as OPCUA or MQTT). The data packet uploaded by the edge computing node to the cloud collaborative platform contains fields such as a timestamp, the ID of the energy storage vehicle, location information, charge and discharge status, and battery health status. The upload frequency ranges from once per second (core scheduling area) to once per minute (standby scheduling area). The instruction packet sent by the cloud collaborative platform to the edge computing node and the on-vehicle control terminal contains fields such as instruction type, target device ID, parameter value, and execution time. For emergency instructions in the core scheduling area, a reliable transmission protocol (such as TCP) is used to ensure that the instructions are not lost, and the transmission delay needs to be controlled within 100 milliseconds; for non-emergency instructions, the UDP protocol can be used to improve the transmission efficiency. In addition, to cope with abnormal situations such as communication interruptions, each node has local caching and offline operation capabilities. For example, the edge computing node can store the historical data and preset control strategies of the last 1 hour, and automatically switch to the locally predefined energy scheduling mode when disconnected from the cloud to maintain the basic operation of the system.

[0077] The implementation of the entire heterogeneous hierarchical deployment architecture also needs to consider the matching of hardware selection and software optimization. Edge computing nodes can adopt embedded processors based on the ARM architecture (such as the Nvidia Jetson series), which have both high performance and low power consumption characteristics and are suitable for deployment in vehicle-mounted or distributed scenarios; the cloud collaboration platform is built based on an x86 server cluster, and containerization technology (such as Docker) is used to achieve elastic expansion of services; the vehicle-mounted control terminal adopts an automotive-grade microcontroller (such as Infineon AURIX) to meet the reliability requirements in harsh environments such as high temperature and vibration. At the software level, the real-time operating system of the edge computing node needs to perform kernel trimming to remove unnecessary services to reduce memory occupancy and interrupt latency; the optimization algorithm of the cloud collaboration platform needs to adopt parallel computing technology (such as the Spark distributed computing framework) to shorten the time for large-scale data processing; the control algorithm of the vehicle-mounted control terminal needs to reduce the computational complexity through code optimization and fixed-point processing to ensure real-time operation under limited hardware resources.

[0078] Through the definition of the above energy partition rules and the differential configuration of heterogeneous nodes, this architecture can achieve precise scheduling of the energy of mobile energy storage vehicles at different time scales and spatial ranges: edge computing nodes solve local real-time control problems, the cloud collaboration platform processes global optimization and cross-regional coordination, and the vehicle-mounted control terminal ensures the stability of the local system. The three cooperate with each other to form a hierarchical distributed energy management system, effectively improving the response ability and operation efficiency of mobile energy storage vehicles in complex power grid environments.

[0079] Embodiment 2:

[0080] The simultaneous solution process of the dynamic optimization decision model is implemented through a multi-objective iterative solver. The core lies in coupling and solving multi-dimensional constraints such as charge-discharge efficiency, temperature field distribution, and load distribution to generate optimal decision parameters that take into account energy efficiency, thermal safety, and scheduling efficiency. The following details the implementation method in combination with a specific scenario example:

[0081] Suppose that a certain urban power grid enters the peak load period (11:00 - 15:00) at noon in summer, and multiple mobile energy storage vehicles in the area need to cooperate to participate in power grid frequency modulation. At this time, the dynamic optimization decision model needs to solve the collaborative constraint relationship of the charge-discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector according to the real-time collected charge-discharge rate of the battery pack (such as the current charging power of a certain vehicle is 50 kW), ambient temperature distribution (the average temperature of the battery compartment is 28 °C, and the local hot spot reaches 32 °C), and power grid load fluctuation (the current regional load exceeds 110% of the rated capacity).

[0082] Loading of the charge-discharge efficiency matrix and update of capacity attenuation:

[0083] The charge-discharge efficiency matrix stores the energy conversion efficiency parameters at different charge-discharge powers. For example, when the charging power ≤ 30kW, the efficiency is 95%; when the power is in the range of 30 - 60kW, the efficiency linearly decreases to 88%; when it exceeds 60kW, it enters the high-loss range and the efficiency drops to 80%. These matrix parameters are pre-calibrated through battery charge-discharge experiments and stored in the database of the energy efficiency optimization module. During the solution process, first, the charge-discharge powers of each vehicle are mapped to specific entries in the matrix. For example, the above-mentioned 50kW charging power corresponds to an efficiency value of 88%. At the same time, the capacity attenuation factor is updated based on the State of Health (SOH) model of the battery, which calculates the capacity attenuation rate according to data such as the cumulative charge-discharge cycles of the battery and the historical working temperature. For example, if a battery pack has completed 500 cycles and the current SOH is 85%, the capacity attenuation factor is set to 0.85, which is used to correct the available capacity parameter in the energy conservation equation set, enabling the model to reflect the impact of battery aging on the charge-discharge efficiency.

[0084] Embedding of the temperature compensation coefficient tensor and calculation of heat conduction:

[0085] The temperature compensation coefficient tensor is used to describe the influence of temperature on battery performance, including charge-discharge efficiency correction factors, internal resistance change factors, etc. in different temperature ranges. For example, when the battery temperature is between 25 - 30°C, the charge-discharge efficiency correction factor is 1.0 (no compensation); for every 1°C increase in temperature, the charging efficiency correction factor decreases by 0.5%, and the discharging efficiency correction factor decreases by 0.3%. After embedding this tensor into the thermodynamic diffusion equation, the anisotropic heat conduction coefficient needs to be calculated through Fourier's law. Taking the temperature distribution in the battery compartment as an example, assuming the temperature gradient along the length direction (x-axis) of the battery pack is 2°C / m, the width direction (y-axis) is 1.5°C / m, and the height direction (z-axis) is 3°C / m, then the anisotropic heat conduction coefficients on the x, y, and z axes are 0.2W / (m·K), 0.15W / (m·K), and 0.25W / (m·K) respectively. By calculating the heat flux density in different directions, the hot spot area of the battery pack (such as the local hot spot at 32°C mentioned above) can be determined, and then the temperature compensation coefficient can be adjusted. For example, a -5% charge-discharge power penalty is imposed on the battery cells in the hot spot area to inhibit further temperature increase.

[0086] Discretization of the load distribution vector and construction of the task constraint matrix:

[0087] The load distribution vector corresponds to the energy dispatch task issued by the power grid dispatch center. For example, within 1 hour, a certain area requires the energy storage vehicle to provide a total discharge capacity of 200 MWh. This vector is discretized into a time series task queue, with a 15-minute time slice, decomposed into 4 task units (each unit needs to provide 50 MWh). Each task unit contains parameters such as start time, target power, and duration. For example, in the first time slice (11:00 - 11:15), a certain vehicle is required to discharge at a power of 60 kW. A task constraint matrix is constructed in combination with the dynamic priority scheduling strategy, and the priority is determined by factors such as the task urgency, the remaining capacity of the energy storage vehicle, and the location distance. For example, an energy storage vehicle within 5 kilometers of the load center and with a remaining capacity ≥ 80% is given a high priority and needs to execute the discharge task first; vehicles with a distance exceeding 10 kilometers or a remaining capacity < 30% have a lower priority and can be used as backup forces. The task constraint matrix also needs to consider the physical constraints of the energy storage vehicle, such as the maximum charge and discharge power (the maximum discharge power of a certain vehicle is 70 kW), the upper and lower limits of SOC (20% - 90%), etc., to ensure that the assigned tasks are within the vehicle's capabilities.

[0088] The iterative optimization process of the parallel gradient descent algorithm:

[0089] Taking the initial optimization input as an example, assume that the charge-discharge efficiency matrix at the current iteration step is [0.95, 0.88, 0.80] (corresponding to low, medium, and high power intervals), the temperature compensation coefficient tensor is [1.0, 0.995, 0.99] (corresponding to 25°C, 26°C, 27°C), and the load distribution vector is [60 kW, 55 kW, 50 kW, 45 kW] (the discharge power of 4 time slices). The energy residual norm is calculated through the energy efficiency optimization module, which is the square root of the sum of the squares of the difference between the actual energy output and the task demand. Assume that the demand for the first time slice is 50 MWh (the energy corresponding to a power of 60 kW for 15 minutes is 15 kWh × 4 = 60 kWh = 0.06 MWh, which is significantly insufficient. This is only a logical example here). The residual norm reflects this gap and drives the model to adjust the power distribution. The thermal management module calculates the thermal balance residual norm. If the deviation between the currently calculated battery temperature field and the preset safe temperature range (20 - 35°C) exceeds the threshold (such as the average temperature of 28°C is close to the upper limit), a temperature correction signal is generated. The task scheduling module calculates the scheduling delay residual norm. If a vehicle's task execution is delayed by more than 5 minutes due to path planning problems, the priority of its subsequent tasks needs to be adjusted.

[0090] The dynamic weighted geometric mean of three types of residual norms is used to adaptively adjust the learning rate decay coefficient. For example, if the energy efficiency residual accounts for 60%, the thermal residual accounts for 30%, and the scheduling residual accounts for 10%, and the current residual norm convergence rate is relatively fast (e.g., the residual decrease amplitude between two adjacent iterations > 10%), then the learning rate amplification mechanism is triggered, multiplying the base learning rate (e.g., 0.1) by the amplification factor (e.g., 1.2) to obtain the new learning rate (0.12) to accelerate the optimization process. If the convergence rate is gentle (the decrease amplitude < 5%), then the decay factor (e.g., 0.9) is used to reduce the learning rate (0.09) to avoid over-optimization. When the residual norm shows periodic fluctuations (e.g., the residual first decreases and then increases for three consecutive iterations), it indicates being trapped in a local optimum. At this time, the random restart strategy is enabled, randomly adjusting the initial values of the charge-discharge power by ±10% and the temperature compensation coefficient by ±5% to re-search for the optimal solution.

[0091] During the iteration process, each module exchanges information in real time through the state sharing interface. For example, when the energy efficiency optimization module finds that the efficiency of a certain vehicle drops sharply during high-power discharge, it transmits the charge-discharge efficiency correction amount (e.g., -8%) to the thermal management module. The thermal management module adjusts the temperature threshold of the vehicle accordingly (e.g., reducing the safety upper limit from 35°C to 32°C) and sends a power limit signal to the task scheduling module. The task scheduling module then reallocates the tasks of this vehicle, transferring part of the load to other vehicles with higher efficiency. Through this cross-module coupling mechanism, the model can comprehensively consider multi-dimensional constraints and avoid system imbalance caused by single-objective optimization.

[0092] After several iterations (e.g., 10 - 20 times), when the dynamic weighted geometric mean is lower than the preset optimization threshold (e.g., 0.05), the iteration is terminated and the optimized decision parameters are output. For example, the charge-discharge efficiency matrix is adjusted to [0.95, 0.89, 0.82] (the efficiency in the medium-power range is increased by 1%), the temperature compensation coefficient tensor is updated to [1.0, 0.996, 0.992] (the efficiency loss per 1°C increase in temperature is reduced), and the load distribution vector is optimized to [65kW, 60kW, 55kW, 40kW] (the power is increased in the first two time slices to cope with peak loads, and the power is reduced in the last two time slices to avoid battery overload). These parameters are sent to the on-vehicle control terminal through the edge computing node to guide the energy storage vehicle to perform specific charge-discharge operations.

[0093] The entire solution process needs to be completed within several minutes to adapt to the rapid change of the grid load. Through the coordinated operation of the multi-objective iterative solver, the dynamic optimization decision model can find a balance among energy efficiency, thermal safety, and task scheduling, ensuring that the mobile energy storage vehicle can meet the load demand while participating in grid frequency modulation, and avoiding battery overheating or excessive loss, thus achieving the dynamic optimum of energy management.

[0094] Example 3:

[0095] The multi-objective iterative solver, as the core execution unit of the dynamic optimization decision-making model, consists of an energy efficiency optimization module, a thermal management module, and a task scheduling module. Each module realizes cross-module coupling through a state sharing interface, forming a closed-loop system for collaborative optimization with multi-dimensional constraints. The following describes the implementation details in combination with the specific operation logic:

[0096] I. Module Function Definition and Basic Configuration

[0097] The energy efficiency optimization module integrates a battery life prediction model. This model constructs a life prediction equation based on parameters such as the number of battery charge-discharge cycles, operating temperature, depth of discharge (DOD), etc. through a linear regression algorithm. For example, the input parameters of the model include: the current number of cycles , the average operating temperature , and the average depth of discharge , and the output is the remaining battery cycle life , and its simplified expression is:

[0098]

[0099] Among them, is the fitting function, which is obtained by training with historical battery test data. The module calculates the impact of different charge-discharge strategies on battery life according to this model. For example, deep discharge (DOD > 80%) or high-temperature environment ( ) will significantly reduce , thus introducing a life loss cost item into the optimization objective to suppress short-sighted high-power charge-discharge behaviors.

[0100] The thermal management module builds a phase change material thermal buffer model, which simulates the heat absorption and release characteristics of phase change materials during the melting / solidification process. Taking paraffin-based phase change materials as an example, its phase change temperature range is 28 - 32 °C, and the latent heat is 200 kJ / kg. The input parameters of the model include: the mass of the phase change material , the current temperature , and the heat flux density , and the output is the phase change progress of the material and the temperature change rate . When the battery pack temperature rises to the lower limit of the phase change temperature (28 °C), the phase change material begins to absorb heat and melt, and the absorbed heat ( is the latent heat) can offset part of the battery heat generation and delay the temperature rise rate; when the temperature drops below 28 °C, the material solidifies and releases heat to maintain the temperature stability of the battery pack. The module calculates the temperature evolution trend under different charge-discharge powers through this model, providing a thermal constraint boundary for energy efficiency optimization.

[0101] The task scheduling module adopts a real-time preemptive scheduling algorithm, which assigns priorities to each energy scheduling task , the priority is determined by factors such as the urgency of the task , delay tolerance , the distance of the energy storage vehicle and so on. For example, for an emergency frequency regulation task (such as when the grid frequency deviation exceeds ±0.5Hz), the is set to the highest level (level 10), and the delay tolerance is set to 5 minutes; for a non-emergency valley period charging task is level 1, and is set to 1 hour. When the algorithm receives a new task, it immediately compares its priority with the priority of the currently executing task. If the new task has a higher priority, the current task is paused and switched to the new task for execution to ensure the real-time performance of high-priority tasks.

[0102] II. Cross-module coupling mechanism and state transfer

[0103] The state sharing interface serves as a channel for data interaction between modules and transfers three types of correction quantities during each iteration:

[0104] Charge and discharge efficiency correction quantity : Calculated by the energy efficiency optimization module, reflecting the impact of the current charge and discharge strategy on the battery life. For example, when a vehicle discharges at a power of 70kW, the battery life prediction model shows that the remaining cycle life is reduced by 100 times. Based on this, the module generates (indicating a virtual 5% reduction in efficiency to suppress this power level) and transfers it to the thermal management module and the task scheduling module. After receiving it, the thermal management module uses this correction quantity as the basis for adjusting the temperature boundary conditions. For example, it lowers the temperature safety upper limit of this vehicle from 35°C to 32°C; the task scheduling module reduces the task priority at this power to avoid frequent use of high-loss operating conditions.

[0105] Temperature gradient correction quantity : Calculated by the thermal management module based on the temperature field distribution inside the battery compartment. For example, when the temperature gradient of a vehicle's battery pack in the x-axis direction reaches 3°C / m (exceeding the safety threshold of 2°C / m), the module generates (dimensionless, representing the degree of gradient overrun) and transfers it to the energy efficiency optimization module and the task scheduling module. The energy efficiency optimization module adjusts the charge and discharge efficiency matrix according to this correction quantity and applies a power penalty to the area with a high temperature gradient (such as reducing the charge and discharge power of the battery cells in the corresponding area by 10%); the task scheduling module preferentially assigns low-power tasks (such as 20kW charging) to this vehicle to reduce heat generation and balance the temperature field.

[0106] Task priority correction quantity : Calculated by the task scheduling module based on the matching degree between the task execution status and the vehicle's capabilities. For example, when the remaining capacity of a vehicle is only 25% (lower than the safety lower limit of 30%), and the currently assigned discharge task requires 15% of the capacity, the module generates (Priority decreased by 3 levels), and is transmitted to the energy efficiency optimization module and the thermal management module. Based on this, the energy efficiency optimization module restricts the upper limit of the vehicle's discharge power (e.g., from 60 kW to 40 kW) and triggers a charging recommendation; the thermal management module adjusts its temperature compensation coefficient to allow a slightly higher operating temperature (e.g., the upper limit rises to 38°C) to prioritize the completion of the remaining tasks and avoid task interruption due to insufficient capacity.

[0107] III. Specific Process of Module Collaborative Optimization

[0108] Taking the example of a mobile energy storage vehicle cluster participating in the power grid peak shaving scenario:

[0109] Initial parameter input: The cloud collaborative platform issues a task queue, requiring that during the period from 14:00 to 15:00, the cluster needs to provide a total discharge capacity of 150 MWh, while ensuring that the battery life loss is less than 5 cycle equivalent values, and the average temperature of the battery pack does not exceed 35°C. The energy efficiency optimization module calculates the initial charge-discharge power distribution plan (e.g., 6 vehicles each undertake 25 MWh, with a power of 50 kW for 1 hour) based on the SOH of each vehicle (e.g., average SOH = 88%) and the historical cycle times (average N = 450 times).

[0110] Cross-module interaction and iteration:

[0111] After receiving the power distribution plan, the thermal management module calculates the temperature evolution of each vehicle's battery pack. Assuming that the battery compartment of vehicle A has poor ventilation and the calculated average temperature will reach 37°C (exceeding the threshold by 2°C) after 1 hour, it generates and transmits it to other modules. After the energy efficiency optimization module responds, it corrects the charge-discharge efficiency of vehicle A to (original value 88%), and the task scheduling module reduces its task priority from level 5 to level 3, triggering task reallocation (e.g., vehicle B undertakes the 5 MWh task of vehicle A, and the power is increased to 60 kW).

[0112] After the task scheduling module reallocates the tasks, it needs to verify whether the remaining capacity of each vehicle meets the new task requirements. Assuming that the remaining capacity of vehicle B is 75%, after undertaking the 5 MWh task, the capacity will drop to 75% - (5 / battery capacity) (assuming the battery capacity is 50 kWh, then it drops to 65%), meeting the constraint of ≥30%. However, the energy efficiency optimization module calculates that when vehicle B discharges at a power of 60 kW, the cycle life loss will increase by 1.2 times (exceeding the single-task allowable upper limit of 0.8 times), generating and transmits it to the thermal management module and the task scheduling module. Based on this, the thermal management module requires that the discharge power of vehicle B does not exceed 55 kW, and the task scheduling module extends its task execution time to 1.1 hours (power 55 kW, capacity 55×1.1 = 60.5 kWh ≈ 6.05 MWh, slightly exceeding the demand, adjusted to 1.09 hours by rounding).

[0113] After 3 iterations, the dynamically weighted geometric mean of the three types of residual norms (energy efficiency residual, thermal residual, scheduling residual) is lower than the preset threshold, and the final parameters are output by each module: the power of vehicle A is adjusted to 40 kW (undertaking 20 MWh, taking 0.5 hours + subsequent charging supplement), the power of vehicle B is 55 kW (undertaking 30 MWh, taking 1.09 hours), the power of the remaining vehicles remains 50 kW, the total task completion time is controlled within 1 hour, and the life and temperature constraints are satisfied simultaneously.

[0114] State feedback and closed-loop control: The optimized decision parameters are sent to the vehicle-mounted control terminal through the edge computing node, and each vehicle performs real-time charge and discharge operations. Vehicle-mounted sensors continuously collect data such as voltage, current, and temperature, and feedback to the multi-objective iterative solver to form a closed loop of "data collection - optimization calculation - instruction execution - state feedback". For example, when vehicle B discharges at 55 kW, the real-time temperature monitoring shows that the average temperature is 34°C and the temperature gradient is 1.8°C / m, both within the safe range and no further adjustment is required; during the 40 kW charging process of vehicle A, the temperature drops to 29°C, and the energy efficiency optimization module gradually restores its charge and discharge efficiency to the original value of 88% according to the feedback data, and reallocates a reasonable load in subsequent tasks.

[0115] IV. Key Points of Hardware and Algorithm Adaptation

[0116] The real-time performance of the multi-objective iterative solver depends on hardware acceleration and algorithm optimization:

[0117] The battery life prediction model of the energy efficiency optimization module uses the look-up table method for acceleration. The life loss parameters under different working conditions are pre-stored as a two-dimensional table (the horizontal axis is DOD, and the vertical axis is temperature), and the results are quickly obtained through linear interpolation during operation, avoiding complex calculation time consumption.

[0118] The phase change material model of the thermal management module uses a simplified lumped parameter method, regarding the battery pack and the phase change material as a single thermal node, ignoring the spatial distribution details, and the calculation speed is increased by more than 50%, meeting the millisecond-level iteration requirements.

[0119] The real-time preemptive algorithm of the task scheduling module is implemented through the hardware interrupt mechanism. When a high-priority task arrives, the CPU immediately responds to the interrupt, saves the current task context (register status, memory pointer, etc.) and switches to the new task, and the interrupt response time is controlled within 10 microseconds to ensure the real-time performance of task scheduling.

[0120] Through the above module design and cooperation mechanism, the multi-objective iterative solver can achieve dynamic optimization of the energy management of mobile energy storage vehicles under complex constraint conditions, taking into account the grid scheduling requirements, battery health status and thermal safety boundaries, providing core technical support for the efficient and reliable operation of mobile energy storage systems.

[0121] Example 4:

[0122] In the actual operation scenario of a certain urban power grid, the phased learning rate adjustment strategy of the parallel gradient descent algorithm realizes the dynamic optimization of the energy management of mobile energy storage vehicles through the adaptive momentum mechanism. When the grid load reaches its peak in the afternoon of summer, multiple mobile energy storage vehicles in the area need to cooperate in peak shaving. In the initial stage, the system takes the charge-discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector of each vehicle as the optimization inputs. For example, the current charging power of a certain vehicle is 50 kW, corresponding to the 88% efficiency value in the charge-discharge efficiency matrix; the average temperature of the battery compartment is 28 °C, and the correction factor corresponding to this temperature in the temperature compensation coefficient tensor is 0.995; in the load distribution vector, the task of this vehicle during the period from 14:00 to 14:15 is to deliver 60 kW of electricity to the grid.

[0123] The energy efficiency optimization module first calculates the energy residual norm, that is, the deviation between the actual energy output and the task demand. If the actual discharge amount of this vehicle during this period is only 55 kW (due to the increase in battery internal resistance resulting in a decrease in efficiency), then the energy residual is 5 kW. The thermal management module synchronously calculates the thermal balance residual norm, and obtains the temperature values of each monitoring point through the temperature sensor network arranged in the battery compartment. Assuming that the temperature of a certain monitoring point is 32 °C, exceeding the safety threshold of 30 °C, a thermal residual signal is generated. The task scheduling module calculates the scheduling delay residual norm. If this vehicle fails to reach the designated area on time due to traffic congestion, resulting in a 3-minute delay in task execution, a scheduling residual is triggered.

[0124] According to the dynamic weighted geometric mean of the three types of residual norms, the system adaptively adjusts the learning rate decay coefficient of the next iteration step. If in the current iteration step, the energy residual accounts for 60%, the thermal residual accounts for 30%, the scheduling residual accounts for 10%, and the convergence rate of the residual norm is higher than the preset threshold (such as the decrease amplitude of the residual between two adjacent iterations > 10%), it indicates that the optimization process is progressing smoothly. At this time, the system multiplies the learning rate amplification factor by the reference learning rate to obtain a new learning rate. For example, if the reference learning rate is 0.1 and the amplification factor is 1.2, the new learning rate is adjusted to 0.12 to accelerate convergence.

[0125] After the learning rate is adjusted, the system recalculates each optimization parameter. The energy efficiency optimization module adjusts the charge-discharge power distribution according to the new learning rate, and increases the task power of this vehicle during the period from 14:15 to 14:30 to 65 kW to make up for the energy gap in the previous period. The thermal management module synchronously adjusts the cooling strategy, increases the fan speed, and controls the battery compartment temperature within the safe range. The task scheduling module replans the vehicle path, avoids congested roads, and ensures that subsequent tasks are executed on time.

[0126] As the iteration progresses, if the convergence rate of the residual norm enters a flat region (the decrease amplitude < 5%), it indicates that the optimization process enters the fine-tuning stage. At this time, the system multiplies the learning rate decay factor by the base learning rate to reduce the learning rate. For example, if the decay factor is 0.9 and the base learning rate is 0.1, the new learning rate is adjusted to 0.09 to improve the optimization accuracy and avoid the algorithm from oscillating near the optimal solution.

[0127] In some cases, the iteration process may fall into a local optimum, manifested as the residual norm showing periodic fluctuations. For example, in three consecutive iterations, the residual norm first decreases and then increases. At this time, the system enables the random restart strategy, randomly adjusting the initial values of the charge and discharge power by ±10% and the temperature compensation coefficient by ±5%. For example, the charge and discharge power of the vehicle is randomly adjusted to 55kW, and the temperature compensation coefficient is adjusted to 0.99, and then re-search for the optimal solution to jump out of the local optimum trap.

[0128] During the entire optimization process, each module exchanges information in real time through the state sharing interface. When the energy efficiency optimization module finds that increasing the power will cause the battery temperature to rise, it immediately transmits the charge and discharge efficiency correction amount to the thermal management module. The thermal management module adjusts the temperature threshold accordingly and sends a power limit signal to the task scheduling module. The task scheduling module then reallocates tasks to ensure that the system maintains a balance among energy efficiency, thermal safety, and scheduling efficiency.

[0129] After multiple iterations, when the dynamic weighted geometric mean is lower than the preset optimization threshold, the system determines that the iteration process converges, terminates the iteration, and outputs the optimized decision parameters. For example, the charge and discharge efficiency matrix is adjusted to [0.95, 0.89, 0.82], the temperature compensation coefficient tensor is updated to [1.0, 0.996, 0.992], and the load distribution vector is optimized to [65kW, 60kW, 55kW, 40kW]. These parameters are sent to the on-vehicle control terminal through the edge computing node to guide the energy storage vehicle to perform specific charge and discharge operations.

[0130] In actual operation, the system continuously repeats the above optimization process according to real-time data. For example, when the grid load fluctuates, the system will recalculate the energy residual norm; when the ambient temperature changes, the thermal management module will update the thermal balance residual norm. Through this continuous optimization, the system can adapt to various complex scenarios, ensuring that when the mobile energy storage vehicle participates in grid regulation, it not only meets the load demand but also guarantees the safety and life of the battery.

[0131] The entire optimization process requires no manual intervention and is fully automated. From data collection to decision output, the entire cycle is controlled within seconds to minutes, meeting the requirements of real-time grid scheduling. Through the phased learning rate adjustment strategy of the parallel gradient descent algorithm, the system can significantly improve the convergence speed while ensuring the optimization accuracy, providing efficient and reliable decision support for the energy management of the mobile energy storage vehicle.

[0132] Example 5:

[0133] The three - level adaptive response instruction sequence is generated by an event - driven state machine, which realizes hierarchical monitoring of the mobile energy storage vehicle and the power grid system based on multi - dimensional trigger conditions. The implementation method is described in detail below in combination with specific scenarios:

[0134] I. Local battery overload warning (primary trigger condition)

[0135] A certain mobile energy storage vehicle performs a rapid discharge task during the peak grid load period. The on - vehicle current sensor continuously collects the output current data of the battery pack. The system extracts the mean value, range, and kurtosis characteristics of the current monitoring data within a fixed time window (such as 100 milliseconds). For example, the 5 current sample values collected at a certain moment are 58A, 62A, 65A, 59A, and 61A. The calculated mean value is 61A, the range is 7A (the difference between the maximum value 65A and the minimum value 58A), and the kurtosis value is 3.2 (reflecting the peak degree of current fluctuation).

[0136] The probability distribution function of the current peak value is fitted through the generalized Pareto distribution model, which is used to describe the probability distribution of extreme current events exceeding a certain threshold. Assuming that the current threshold in historical data is set to 60A, the model determines the distribution parameters through the maximum likelihood estimation method, and calculates that the upper limit of the 99.9% confidence interval is 68A. When the measured current value at a certain moment reaches 70A (exceeding the upper limit of the confidence interval), the system determines that a current mutation has occurred and triggers a local battery overload warning instruction.

[0137] After the warning is triggered, the event - driven state machine executes nested conditional jump logic: First, it sends a power reduction instruction to the on - vehicle control terminal to reduce the discharge power from 60kW to 40kW to reduce the current; at the same time, it starts a sliding time window extension mechanism to extend the monitoring window from 100 milliseconds to 500 milliseconds to continuously observe the current fluctuation trend. If subsequent monitoring shows that the current drops below 60A and remains stable, the current power is maintained; if the current continues to be higher than the threshold, it further triggers the strong start instruction of the battery cooling system, raises the fan speed to the highest gear, and reports the abnormal state to the edge computing node to request adjustment of the task allocation of this vehicle.

[0138] II. Regional power grid frequency modulation warning (secondary trigger condition)

[0139] In the power grid of an industrial park, the system collects the power grid frequency fluctuation signal (such as the rated frequency of 50Hz) through monitoring equipment deployed in the substation. The multi - scale frequency - domain analysis algorithm uses wavelet transform decomposition technology to map the frequency signal into a multi - resolution wavelet coefficient matrix. For example, the frequency signal within 10 seconds is decomposed by wavelet three - layer, and the low - frequency approximation component (reflecting the fundamental frequency trend) and high - frequency detail components (reflecting the harmonic components) are obtained.

[0140] By extracting the high-frequency abnormal components in the matrix, it was found that in a certain period (14:20-14:25), there was a coefficient with a significantly increased amplitude in the 250Hz frequency band (such as the coefficient value jumped from 0.5 to 3.0), indicating the presence of 5th harmonic interference (250Hz=5×50Hz). Based on the cluster analysis algorithm, the high-amplitude coefficient clusters associated in time and space were merged to generate potential harmonic interference area markers (such as Area A in the industrial park). The energy proportion (the proportion of harmonic energy to total energy) and the frequency offset product index (frequency deviation × duration) of each marked area were calculated. Assuming that the energy proportion of Area A reached 15%, the frequency offset product was 0.3Hz·s (exceeding the preset frequency regulation level threshold of 0.2Hz·s), and the system triggered the regional power grid frequency regulation warning.

[0141] After the warning is generated, the state machine performs a hierarchical response: first, it sends the harmonic interference area location information to the cloud collaboration platform, requesting to call the mobile energy storage vehicle in the area to perform the harmonic suppression task; at the same time, it sends instructions to the relevant on-board control terminal to adjust the inverter control parameters of the energy storage vehicle, such as increasing the PWM modulation frequency from 10kHz to 15kHz to suppress the generation of the 5th harmonic. If the harmonic interference lasts for more than 10 minutes, the energy storage vehicle in the standby dispatch area will be further started to move to the area to enhance the frequency regulation capability of the local power grid.

[0142] 3. Prediction of global energy scheduling imbalance (ultimate trigger condition)

[0143] In the cross-regional grid interconnection scenario, the system collects load data and energy status of energy storage vehicles in each region in real time. The energy distribution entropy value is used to measure the uniformity of the system's energy distribution. The higher the entropy value, the more disordered the distribution and the greater the risk of instability. For example, at a certain moment, the energy distribution entropy value of the entire network is calculated to be 2.8 (the preset threshold is 2.5), indicating that the energy distribution tends to be dispersed, some areas (such as the downtown commercial area) are overloaded, and the energy idle rate of suburban energy storage stations exceeds 40%.

[0144] When the entropy value exceeds the threshold, the system determines that there is a risk of global energy scheduling imbalance and triggers the final warning. The event-driven state machine first performs energy distribution entropy value tracing analysis to locate the main contributing areas of increased entropy value (such as commercial area load exceeding the rated capacity by 120%, and suburban energy storage vehicle SOC reaching an average of 90%). Then a global energy scheduling reorganization instruction is generated: a batch charging instruction is issued to the mobile energy storage vehicles at the suburban energy storage station to replenish energy using low-priced electricity during off-peak hours; at the same time, an emergency discharge instruction is sent to the energy storage vehicles around the commercial area to increase the charging and discharging power to 90% of the rated value, and the traffic management system is coordinated to open a green passage for the energy storage vehicle to shorten its time to reach the load center.

[0145] To avoid system oscillations, the state machine sets a warning delay mechanism: if the entropy value drops back to the safe range within 15 minutes after triggering the threshold, the reorganization instruction is cancelled; if it continues to rise, energy storage vehicles in the standby dispatching area are activated for cross-regional support. For example, 10 energy storage vehicles are dispatched from a standby site 50 kilometers away to the commercial area and are expected to arrive and participate in the dispatching within 2 hours.

[0146] IV. Nested Logic of the State Machine and Instruction Coordination

[0147] The nested conditional jump logic of the event-driven state machine is implemented through multi-level state nodes. For example, under the local battery overload warning state node, there are sub-nodes such as "current returns to normal", "cooling system starts", and "task reallocation". Corresponding jumps are executed according to different feedback signals:

[0148] If the current drops below the threshold within 5 minutes after reducing the power, jump to the "normal monitoring" state;

[0149] If the temperature continues to rise after the cooling system starts, jump to the "emergency shutdown" state, cut off the charge and discharge circuit, and report the fault;

[0150] If the current is still over-limit after task reallocation, jump to the "exit dispatching" state, mark the vehicle as to be repaired, and remove it from the task queue.

[0151] In the collaborative scenario of regional power grid frequency modulation warning and global imbalance prediction, when a harmonic warning in a certain region is triggered, if the cloud collaborative platform fails to effectively suppress the harmonics within 30 minutes (such as the harmonic energy ratio rises back to 20%), the state machine automatically raises the response level, triggers the global imbalance prediction process, incorporates the frequency modulation demand of this region into the whole-network energy dispatching optimization model, and achieves a higher level of balance through cross-regional energy allocation.

[0152] V. Hardware Deployment and Real-time Guarantee

[0153] Local current monitoring relies on high-precision Hall current sensors (accuracy ±0.5%), with a sampling frequency of 10 kHz to ensure capturing instantaneous current peaks; grid frequency analysis uses a dedicated signal acquisition card (such as NIPXIe-5124), which supports synchronous acquisition of multiple signals and real-time wavelet transform; the calculation of the energy distribution entropy value is completed by a cloud server cluster, and the calculation delay is controlled within 2 seconds using GPU acceleration technology.

[0154] The event-driven state machine realizes microsecond-level task scheduling through a real-time operating system (such as RTX) to ensure the real-time nature of trigger condition detection and instruction generation. For example, the delay of current mutation detection does not exceed 50 microseconds, the frequency offset calculation delay does not exceed 10 milliseconds, and the global entropy value calculation delay does not exceed 1 second, meeting the requirements of real-time monitoring and rapid response of the power grid.

[0155] Through the above three-level response mechanism and state machine logic, the system can conduct hierarchical early warning and intervention from local device anomalies to global system risks, forming a complete closed-loop covering "detection - early warning - response - adjustment", effectively improving the reliability of the energy management of mobile energy storage vehicles and the stability of power grid operation.

[0156] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0157] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for energy management of a mobile energy storage vehicle based on an intelligent algorithm, characterized in that, The method includes: Collecting multi-dimensional energy state data of a mobile energy storage vehicle, where the multi-dimensional energy state data includes the real-time charge and discharge rate of the battery pack, the ambient temperature distribution characteristics, and the grid load fluctuation characteristics; Generating a heterogeneous hierarchical deployment architecture for edge computing nodes, cloud collaborative platforms, and vehicle-mounted control terminals based on the dynamic response requirements of the energy scheduling unit; Synchronously obtaining the real-time energy interaction data streams of each node under the heterogeneous hierarchical deployment architecture, and performing multi-source heterogeneous data fusion and outlier rejection processing on the real-time energy interaction data streams; Constructing a dynamic optimization decision model through a battery degradation model and a grid supply-demand balance equation, and jointly solving the collaborative constraint relationships of the charge and discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector in the dynamic optimization decision model; Generating a three-level adaptive response instruction sequence including local battery overload alarms, regional grid frequency modulation warnings, and global energy scheduling imbalance predictions based on the boundary trigger conditions of the collaborative constraint relationships; 2. The energy management method of the mobile energy storage vehicle based on intelligent algorithm according to claim 1, characterized in that, The heterogeneous hierarchical deployment architecture is characterized by a distributed computing layer; The generating of the heterogeneous hierarchical deployment architecture for edge computing nodes, cloud collaborative platforms, and vehicle-mounted control terminals includes: Defining the energy partition rules for the core scheduling area, buffer scheduling area, and standby scheduling area according to the peak-valley period division strategy in the grid load fluctuation characteristics; Based on the energy partition rules, configuring a low-latency response mode for the edge computing nodes, a global optimization mode for the cloud collaborative platform, and a local closed-loop control mode for the vehicle-mounted control terminals; 3. The energy management method of the mobile energy storage vehicle based on the intelligent algorithm according to claim 2, characterized in that, The dynamic optimization decision model is implemented through a multi-objective iterative solver; the jointly solving the collaborative constraint relationships of the charge and discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector in the dynamic optimization decision model includes: Loading the charge and discharge efficiency matrix into the energy conservation equations of the multi-objective iterative solver, and updating the capacity attenuation factor based on the battery health state model; Embedding the temperature compensation coefficient tensor into the thermodynamic diffusion equation, and calculating the anisotropic thermal conductivity through Fourier's law; Discretizing the load distribution vector into a time-series task queue, and constructing a task constraint matrix in combination with a dynamic priority scheduling strategy; Using a parallel gradient descent algorithm to synchronously update the optimization residuals of the energy conservation equations, thermodynamic diffusion equation, and task constraint matrix until the steady-state convergence condition is reached; 4. The energy management method of the mobile energy storage vehicle based on intelligent algorithm according to claim 3, characterized in that, The multi-objective iterative solver includes an energy efficiency optimization module, a thermal management module, and a task scheduling module; the energy efficiency optimization module, thermal management module, and task scheduling module are cross-module coupled through a state sharing interface; A battery life prediction model is configured in the energy efficiency optimization module, a phase change material thermal buffer model is configured in the thermal management module, and a real-time preemptive scheduling algorithm is configured in the task scheduling module; The state sharing interface transfers the charge and discharge efficiency correction amount, temperature gradient correction amount, and task priority correction amount to adjacent functional modules during each iteration; 5. The energy management method of the mobile energy storage vehicle based on the intelligent algorithm according to claim 4, characterized in that The parallel gradient descent algorithm is implemented using a staged learning rate adjustment strategy; the using a parallel gradient descent algorithm to synchronously update the optimization residuals of the energy conservation equations, thermodynamic diffusion equation, and task constraint matrix includes: Take the charge-discharge efficiency matrix, temperature compensation coefficient tensor, and load distribution vector of the current iteration step as the initial optimization input; Calculate the energy residual norm through the energy efficiency optimization module, calculate the thermal balance residual norm through the thermal management module, and calculate the scheduling delay residual norm through the task scheduling module; Adaptively adjust the learning rate decay coefficient of the next iteration step according to the dynamic weighted geometric mean of the three types of residual norms; Terminate the iteration when the dynamic weighted geometric mean is lower than the preset optimization threshold, and output the optimized decision parameters.

6. The energy management method of the mobile energy storage vehicle based on the intelligent algorithm according to claim 5, characterized in that, The adjustment strategy of the learning rate decay coefficient is implemented based on the adaptive momentum mechanism; the adaptive adjustment of the learning rate decay coefficient of the next iteration step includes: If the convergence rate of the residual norm in the current iteration step is higher than the rate threshold, multiply the learning rate amplification factor by the reference learning rate as the new parameter; If the convergence rate of the residual norm is in a flat interval, multiply the learning rate decay factor by the reference learning rate as the new parameter; If the residual norm shows periodic fluctuations, enable the random restart strategy to reset the optimization direction and step size range.

7. The energy management method of the mobile energy storage vehicle based on intelligent algorithm according to claim 1, characterized in that The three-level adaptive response instruction sequence is generated by an event-driven state machine; the generation of the three-level adaptive response instruction sequence including local battery overload warning, regional power grid frequency modulation warning, and global energy scheduling imbalance prediction includes: Set a primary trigger condition based on current mutation for local battery overload warning, and use the sliding time window statistical method to detect the instantaneous current abnormal peak value; Set a secondary trigger condition based on frequency deviation for regional power grid frequency modulation warning, and identify the power grid harmonic distortion area through the multi-scale frequency domain analysis algorithm; Set a final trigger condition based on load deviation exceeding the limit for global energy scheduling imbalance prediction, and use the energy distribution entropy value calculation to determine the instability critical point of the system.

8. The energy management method of the mobile energy storage vehicle based on the intelligent algorithm according to claim 7, characterized in that The event-driven state machine contains nested conditional jump logic; the use of the sliding time window statistical method to detect the instantaneous current abnormal peak value includes: Extract the mean, range, and kurtosis features of the current monitoring data within a fixed time window; Fit the current peak probability distribution function through the generalized Pareto distribution model; When the observed current value exceeds the upper limit of the 99.9% confidence interval of the probability distribution function, trigger the local battery overload warning instruction.

9. The energy management method of the mobile energy storage vehicle based on the intelligent algorithm according to claim 8, characterized in that, The multi-scale frequency domain analysis algorithm is implemented by using wavelet transform decomposition technology; The identification of the power grid harmonic distortion area through the multi-scale frequency domain analysis algorithm includes: Map the power grid frequency fluctuation signal into a multi-resolution wavelet coefficient matrix, and extract the high-frequency abnormal components in the matrix; Based on the clustering analysis algorithm, merge the high-amplitude coefficient clusters with spatio-temporal correlation to generate potential harmonic interference area markers; Calculate the energy proportion and frequency deviation product index of each marked area, and screen the target intervals that meet the preset frequency modulation level.

10. The energy management method of the mobile energy storage vehicle based on intelligent algorithm according to claim 1, characterized in that The multi-source heterogeneous data fusion and outlier rejection processing are implemented by using a robust data cleaning strategy; the multi-source heterogeneous data fusion and outlier rejection processing of the real-time energy interaction data stream includes: Establish the data mapping relationship between the current time series sequence of the edge computing node, the load frequency sequence of the cloud platform, and the temperature sampling sequence of the vehicle-mounted terminal; Separate the high-frequency noise components in the data streams of each node through the empirical mode decomposition algorithm, and perform cross-modal alignment on the low-frequency effective data based on the robust principal component analysis algorithm; Use the time series interpolation algorithm to compensate for the sampling interval deviation between multi-source data streams and generate a spatio-temporally consistent fusion energy management tensor.

Images