An ai computing power optimization scheduling system for new energy storage

Abstract

The application relates to the technical field of optimization scheduling, and discloses an AI computing power optimization scheduling system for new energy storage, which comprises the following steps: collecting the instantaneous voltage, current and absolute temperature of a storage node in real time, and inversely solving the electrochemical intrinsic angular frequency; analyzing the directed acyclic graph of an artificial intelligence computing task, and mapping an expected pulse angular frequency; performing time domain interference integration on the two to construct a resonance penalty factor, and combining a three-dimensional space absolute temperature distribution to converge to obtain a maximum computing power bearing extreme boundary; using a high-order derivative as a chaotic signal to drive a particle swarm optimization algorithm to solve an optimal distribution matrix; calculating an execution time phase shift angle and dynamically inserting a null instruction to perform time translation; and finally, packaging heat dissipation into an energy evolution function, and reversely updating a sampling reference frequency of an analog-to-digital converter. The application effectively avoids battery deterioration caused by high-frequency resonance, and significantly improves the safety and operation life of the algorithm storage collaboration.

Description

Technical Field

[0001] This invention relates to the field of optimization scheduling technology, and more specifically, to an AI computing power optimization scheduling system for new energy storage. Background Technology

[0002] With the coordinated development of large-scale artificial intelligence models and distributed new energy storage, integrated computing and storage is gradually becoming an important application scenario. Most existing scheduling strategies are based on macroscopic power matching and average power monitoring, treating energy storage units as static power supply capacitors and allocating tasks solely based on set remaining power limits or surface temperature constraints. However, modern artificial intelligence computing exhibits highly periodic transient pulse power ripple at the operational level, while the ion diffusion and deintercalation processes within the battery inherently exhibit dynamic hysteresis. When the frequency of these high-frequency excitation pulses approaches the relaxation frequency of the cell's dynamic changes, high-frequency resonance at the electromechanical level is easily induced. Existing technologies lack the ability to detect this cross-domain resonance phenomenon, leading to extreme local Joule heating within the cell even when macroscopic data appears normal, due to the sharp amplification of nonlinear polarization impedance. This accelerates battery aging and causes irreversible safety hazards. Existing methods struggle to effectively avoid such risks without interfering with fixed hardware configurations, exhibiting significant safety blind spots and scheduling limitations in complex heterogeneous computing scenarios. Summary of the Invention

[0003] This invention provides an AI computing power optimization and scheduling system for new energy storage, which solves the technical problems mentioned in the background art.

[0004] This invention provides an AI computing power optimization and scheduling system for new energy storage, applied to a system including distributed energy storage nodes and a hardware acceleration unit with a single-instruction multi-threaded architecture and arithmetic logic unit. The system includes an analog-to-digital converter, comprising: Real-time acquisition of instantaneous voltage, instantaneous current and absolute temperature of each node is used to construct a three-dimensional temperature field, and the electrochemical intrinsic angular frequency is calculated in reverse. The directed acyclic graph of multiple tensor computation tasks sent to the hardware acceleration unit is parsed, and the tensor dimension size and dynamic main frequency are extracted and mapped to the expected pulse angular frequency. The resonance penalty factor is constructed by performing time-domain integration between the expected pulse angular frequency and the electrochemical intrinsic angular frequency. Based on the resonance penalty factor and the three-dimensional spatial temperature field, the computing power bearing boundary of each node is converged. The higher-order derivatives of the instantaneous voltage and instantaneous current are extracted as physical perturbations. The optimal allocation matrix is ​​obtained by using the computing power to drive the particle swarm optimization algorithm to find the optimal allocation matrix. Calculate the execution time offset of each task in the optimal allocation matrix, and dynamically insert empty instructions into the underlying drive flow to perform time shifting of the tensor calculation task; The heat dissipation caused by the translation task is encapsulated in the energy evolution function, and its transient derivative is used to update the sampling reference frequency of the analog-to-digital converter in reverse.

[0005] The beneficial effects of this invention are as follows: This solution does not rely on manually preset safety temperature thresholds or remaining power limits. It dynamically converges the computing power bearing capacity boundary by real-time acquisition of electrical signals and temperature data, and actively avoids resonance troughs by utilizing microsecond-level time shifting at the underlying instruction level. This solution significantly reduces the risk of local abnormal heat accumulation and irreversible lithium plating caused by high-frequency load abrupt changes, achieves adaptive coordination between computing power allocation and energy storage response on a multi-dimensional scale, effectively extends the lifecycle health of distributed energy storage networks, and improves the security and stability of concurrent execution of computing tasks. Attached Figure Description

[0006] Figure 1 This is a flowchart of an AI computing power optimization and scheduling system for new energy storage according to the present invention. Detailed Implementation

[0007] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0008] This embodiment operates in a device comprising distributed energy storage nodes and a hardware acceleration unit with a single-instruction multi-threaded architecture and arithmetic logic units. The device is equipped with a multi-channel analog-to-digital converter (ADC). The distributed energy storage nodes are independent electrochemical energy storage units capable of storing and releasing electrical energy. Each distributed energy storage node is equipped with a voltage acquisition module, a current acquisition module, and a multi-point temperature acquisition module. The analog signal output terminals of each acquisition module are connected to the corresponding analog input channels of the ADC, and the digital signal output terminals of the ADC are connected to the system's computational processing unit. The hardware acceleration unit is a computational unit capable of parallel tensor computation. Its computation instruction input terminals are connected to the system's underlying driver module, and its computation status output terminals are connected to the system's computational processing unit.

[0009] like Figure 1As shown, an AI computing power optimization and scheduling system for new energy storage is applied to a system that includes distributed energy storage nodes and a hardware acceleration unit with a single-instruction multi-threaded architecture and an arithmetic logic unit. The system includes an analog-to-digital converter, comprising: Real-time acquisition of instantaneous voltage, instantaneous current and absolute temperature of each node is used to construct a three-dimensional temperature field, and the electrochemical intrinsic angular frequency is calculated in reverse. The directed acyclic graph of multiple tensor computation tasks sent to the hardware acceleration unit is parsed, and the tensor dimension size and dynamic main frequency are extracted and mapped to the expected pulse angular frequency. The resonance penalty factor is constructed by performing time-domain integration between the expected pulse angular frequency and the electrochemical intrinsic angular frequency. Based on the resonance penalty factor and the three-dimensional spatial temperature field, the computing power bearing boundary of each node is converged. The higher-order derivatives of the instantaneous voltage and instantaneous current are extracted as physical perturbations. The optimal allocation matrix is ​​obtained by using the computing power to drive the particle swarm optimization algorithm to find the optimal allocation matrix. Calculate the execution time offset of each task in the optimal allocation matrix, and dynamically insert empty instructions into the underlying drive flow to perform time shifting of the tensor calculation task; The heat dissipation caused by the translation task is encapsulated in the energy evolution function, and its transient derivative is used to update the sampling reference frequency of the analog-to-digital converter in reverse.

[0010] The analog-to-digital converter has a sampling bit depth of no less than 16 bits, the analog input range covers the output signal range of the corresponding acquisition module, the sampling accuracy reference adopts an internal high-precision reference voltage source, the data transmission adopts a serial peripheral interface protocol, the single-channel sampling rate can be configured from 1kHz to 1MHz, and it is updated synchronously with the system's sampling reference frequency.

[0011] The sampling synchronization between each acquisition module and the analog-to-digital converter is achieved using the Network Time Protocol (NTP). The master clock source is deployed in the system's computing unit, and the acquisition modules and hardware acceleration units of each distributed energy storage node act as slave clock nodes, synchronizing with the master clock source. The time synchronization error does not exceed 10% of the sampling period. Sampling triggering adopts an edge-synchronized triggering method. The master clock source sends a synchronization trigger pulse at the beginning of each sampling period. All acquisition modules and the analog-to-digital converter synchronously complete sampling and conversion after receiving the trigger pulse, ensuring that the timestamps of all acquired data are perfectly aligned.

[0012] The underlying driver module runs in the kernel mode of the hardware acceleration unit, supports real-time editing and insertion of the instruction stream of the hardware acceleration unit, is compatible with the native instruction set of the hardware acceleration unit, and uses the native no-operation instruction of the hardware acceleration unit. It can be inserted into any non-dependent interval position of the instruction stream without destroying the data dependency relationship of the tensor calculation task. The number of inserted empty instruction cycles can be configured from 1 to 65535 hardware clock cycles.

[0013] All physical quantities involved in the calculations are first normalized and converted into dimensionless values ​​before proceeding to subsequent calculations. The normalization rules for each physical quantity are as follows: The instantaneous voltage is normalized using the maximum amplitude normalization method, and the calculation formula is as follows: in, The value is the raw instantaneous voltage collected by the distributed energy storage node numbered i at time t, in volts. The rated voltage of the battery cell for distributed energy storage node numbered i is given in volts. The normalized instantaneous voltage is a dimensionless value, ranging from 0 to 1.5.

[0014] The normalization of instantaneous current adopts the maximum amplitude normalization method, and the calculation formula is as follows: in, The value of the raw instantaneous current collected by the distributed energy storage node numbered i at time t is in amperes. The rated discharge current value of the distributed energy storage node numbered i, in amperes; The normalized instantaneous current is dimensionless, taking a positive value in the discharge state and a negative value in the charging state, with a value range of -2 to 2.

[0015] The normalization of absolute temperature uses a linear normalization method, and the calculation formula is as follows: in, The value is the raw absolute temperature collected by the distributed energy storage node numbered i at time t, in Kelvin. This is the lower limit of the operating temperature of the battery cell, and its value is 233.15 Kelvin. This is the upper limit of the cell's operating temperature, set at 333.15 Kelvin. This is the dimensionless value of the normalized absolute temperature, ranging from 0 to 1.

[0016] The normalization of the instantaneous utilization rate of the arithmetic logic unit adopts the linear normalization method, and the calculation formula is as follows: in, The raw instantaneous utilization value of the arithmetic logic unit of the hardware acceleration unit at time t, in percentage form; The instantaneous utilization rate of the normalized arithmetic logic unit is a dimensionless value, ranging from 0 to 1.

[0017] The normalization of the dynamic clock frequency of the hardware acceleration unit adopts the maximum amplitude normalization method, and the calculation formula is as follows: in, The original dynamic clock frequency of the hardware acceleration unit during the runtime of the tensor computation task numbered j, in Hertz; This is the rated maximum clock frequency of the hardware acceleration unit, in Hertz; The value is the dimensionless value of the normalized dynamic main frequency, ranging from 0.1 to 1.

[0018] The spatial coordinates are normalized using a linear normalization method. The three-dimensional spatial coordinates of the distributed energy storage node are: The unit is meters, and the normalized three-dimensional spatial coordinates are... The calculation formula is: in, These represent the maximum and minimum values ​​of the spatial coordinates of all distributed energy storage nodes in their respective dimensions; the normalized spatial coordinates are dimensionless values ​​ranging from 0 to 1.

[0019] The normalization of the time variable uses linear normalization, and the calculation formula is as follows: in, This is the original physical time, in seconds; The initial sampling period of the system is in seconds; This is the normalized dimensionless time variable.

[0020] The timescale matching between the electrochemical process and the tensor calculation process employs a multi-rate sampling and interpolation synchronization method. The basic sampling period of the electrochemical parameters is synchronized with the system's sampling reference frequency, while the sampling period of the tensor calculation parameters is 1 / 1000 of the electrochemical parameter sampling period. A linear interpolation method is used to interpolate the tensor calculation parameters onto the same time axis as the electrochemical parameters, ensuring that the parameters of both processes are matched on the same dimensionless time axis. The interpolation process uses a first-order linear interpolation algorithm. For any dimensionless time t, the corresponding tensor calculation parameter interpolation formula is: in, and For the tensor parameter sampling sequence, the two sampling times adjacent to the target time t are considered. and These are the original values ​​of the tensor calculation parameters at the corresponding sampling time.

[0021] When the relative deviation between the expected pulse angular frequency of the tensor computation task and the electrochemical intrinsic angular frequency of the distributed energy storage node is less than 5% and the duration exceeds 10 dimensionless time units, electrochemical-electrical resonance is determined to have occurred. This resonance leads to nonlinear amplification of the polarization impedance inside the cell, causing local Joule heat accumulation and accelerating cell aging. The value of the resonance penalty factor is positively correlated with the resonance risk; the larger the resonance penalty factor, the higher the resonance risk of the corresponding node and task combination.

[0022] The data acquisition modules configured on each distributed energy storage node synchronously collect the raw instantaneous voltage, raw instantaneous current, and raw absolute temperature of the corresponding node. Dimensionless transformation is performed according to the aforementioned normalization rules. Based on the normalized three-dimensional spatial coordinates and normalized absolute temperature of each distributed energy storage node, a three-dimensional spatial temperature field is constructed. The three-dimensional spatial temperature field is constructed using an inverse distance weighted interpolation algorithm. For any coordinate point in three-dimensional space... The corresponding numerical calculation formula for the temperature field is: Where N is the total number of distributed energy storage nodes; denoted as , where is the dimensionless spatial distance between the target coordinate point and the distributed energy storage node numbered i; p is the inverse distance weighted exponent, with a value of 2; Let t be the dimensionless normalized temperature field value corresponding to the target coordinate point at time t.

[0023] The first derivative of the normalized instantaneous voltage with respect to dimensionless time is multiplied by the normalized instantaneous current, and then the product of the normalized instantaneous voltage and the first derivative of the normalized instantaneous current with respect to dimensionless time is subtracted. The absolute value of the difference is taken as the amplitude of the phase angle change rate. The amplitude of the phase angle change rate characterizes the dynamic rate of phase change of the electrical signal inside the electrochemical energy storage unit, reflecting the dynamic characteristics of ion migration inside the cell. For discretely sampled time series data, the first derivative of the normalized instantaneous voltage with respect to dimensionless time is approximated using the central difference method, and the calculation formula is as follows: in, The dimensionless sampling period is set to 1. and These are the normalized instantaneous voltage values ​​at adjacent times before and after the current sampling time. For the initial sampling time, without a preceding sampling point, the first derivative is calculated using the forward difference method, with the following formula: The first derivative of the normalized instantaneous current with respect to dimensionless time is calculated using the same finite difference method.

[0024] The formula for calculating the magnitude of the rate of change of phase angle is: in, Let be the amplitude of the phase angle change rate corresponding to the distributed energy storage node numbered i at time t, which is a dimensionless value.

[0025] Adding the square of the normalized instantaneous voltage to the square of the normalized instantaneous current yields the sum of squares of the electrical signal amplitude. This sum of squares characterizes the total energy amplitude of the output electrical signal from the distributed energy storage node and is used to normalize the phase change rate, eliminating the influence of overall signal amplitude fluctuations on the calculation results. The formula for calculating the sum of squares of the electrical signal amplitude is: in, Let be the sum of squares of the electrical signal amplitude corresponding to the distributed energy storage node numbered i at time t, which is a dimensionless value. When the calculated... When it is less than 1e-6, take This avoids errors such as division by zero in subsequent calculations.

[0026] Dividing the amplitude of the phase angle change rate by the sum of the squares of the electrical signal amplitudes yields the reference angular frequency component. The reference angular frequency component is the fundamental value of the cell's intrinsic electrochemical oscillation frequency after eliminating the influence of the electrical signal amplitude. It reflects the cell's inherent oscillation characteristics under current operating conditions and is unaffected by the overall output power. The formula for calculating the reference angular frequency component is: in, Let i be the reference angular frequency component corresponding to the distributed energy storage node numbered i at time t, and let i be a dimensionless value.

[0027] The second derivative of normalized absolute temperature with respect to dimensionless time is divided by the normalized absolute temperature, and the quotient is used as the exponent of the natural constant for exponential calculation to obtain the thermally induced modulation term. The thermally induced modulation term characterizes the modulation effect of dynamic temperature changes on the electrochemical oscillation frequency of the battery cell, reflecting the influence of internal thermodynamic processes on ion migration rates. For discretely sampled time-series data, the second derivative of normalized absolute temperature with respect to dimensionless time is approximated using the central difference method, and the calculation formula is as follows: For the initial two sampling times, the second derivative is calculated using the forward difference method, and the formula is as follows: The formula for calculating the thermally excited modulation term is: in, The thermally activated modulation term corresponding to the distributed energy storage node numbered i at time t is a dimensionless value; an offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0028] Multiplying the reference angular frequency component by the thermally excited modulation term yields the electrochemical intrinsic angular frequency of the corresponding distributed energy storage node, calculated using the following formula: in, Let be the electrochemical intrinsic angular frequency of the distributed energy storage node numbered i at time t. It is a dimensionless value and corresponds to the normalized value of the characteristic relaxation angular frequency of the lithium-ion intercalation / deintercalation process inside the cell.

[0029] The system receives and parses a directed acyclic graph (DAG) of multiple tensor computation tasks sent to the hardware acceleration unit, extracting the computational logic and data dependencies for each task. The DAG represents the computational flow topology of the tensor computation tasks; each node corresponds to a tensor operation, and the directed edges represent the data dependencies between operations. A depth-first search algorithm is used to parse the DAG to obtain the complete computational flow of the tensor computation tasks and the tensor parameters of all computation nodes.

[0030] Extract the input tensor, weight tensor, and output tensor of each computation node in the tensor computation task. Select the tensor corresponding to the core computation node with the highest computational cost in the task, and extract the total number of dimensions of this tensor and the size of each dimension. The total number of dimensions of a tensor is its order, and the size of each dimension is the number of elements in that dimension.

[0031] The sum of all dimensions of the tensor corresponding to the core computation node of the tensor computation task yields the tensor dimension sum. The tensor dimension sum represents the total element size of the tensor corresponding to the tensor computation task across all dimensions, reflecting the overall computational scale of the task. The formula for calculating the tensor dimension sum is: Where D is the total number of dimensions of the tensors corresponding to the core operation nodes; Calculate the dimension size of the d-th dimension of the task core tensor for tensor j; The tensor dimension and value are calculated for the tensor task numbered j.

[0032] The geometric mean of the tensor dimensions is obtained by multiplying the sizes of all dimensions of the tensor corresponding to the core operation node of the tensor computation task, and then taking the arithmetic root of the product raised to the power of the total number of dimensions. The geometric mean of the tensor dimensions represents the average element size of the tensor corresponding to the tensor computation task across all dimensions. It is used to eliminate the influence of uneven tensor dimension distribution on the computation results and to normalize the computational characteristics of tensors with different dimensional distributions. The formula for calculating the geometric mean of the tensor dimensions is: in, The geometric mean of the tensor dimension corresponding to the task with tensor number j is calculated. When the calculated... When it is less than 1e-6, take This avoids errors such as division by zero in subsequent calculations.

[0033] The normalized instantaneous utilization rate of the arithmetic logic unit is extracted. The first derivative of the natural logarithm of the normalized instantaneous utilization rate with respect to dimensionless time is calculated, and its absolute value is added to obtain the dynamic load rate derivative term. The dynamic load rate derivative term characterizes the load change rate of the hardware acceleration unit and reflects the dynamic load fluctuation characteristics of the computing task. When the load remains unchanged, the value of the dynamic load rate derivative term is 1; when the load changes, the value of the dynamic load rate derivative term is greater than 1. For discretely sampled time series data, the first derivative of the natural logarithm of the normalized instantaneous utilization rate with respect to dimensionless time is calculated using the central difference method, and the initial time is calculated using the forward difference method. The formula for calculating the dynamic load rate derivative term is: in, The dynamic load rate derivative term corresponding to the tensor computation task numbered j at time t is a dimensionless value; an offset of 1e-6 is added to the logarithmic operation to avoid the error of undefined logarithm.

[0034] The expected pulse angular frequency for the corresponding tensor computation task is obtained by continuously multiplying the normalized dynamic dominant frequency, tensor dimension sum, and dynamic load rate derivative term, and then dividing by the geometric mean of the tensor dimension. The calculation formula is as follows: in, Let be the expected pulse angular frequency of the tensor computation task numbered j at time t. Let be a dimensionless value, corresponding to the normalized value of the characteristic angular frequency of the transient pulse power ripple generated during the operation of the tensor computation task.

[0035] Multiplying the expected pulse angular frequency corresponding to the tensor calculation task at the same dimensionless moment by the electrochemical intrinsic angular frequency corresponding to the distributed energy storage node yields the frequency interference product term. The frequency interference product term characterizes the coupling strength between the two frequencies; the larger the value, the higher the potential resonance risk between the two frequencies. The formula for calculating the frequency interference product term is: in, Let be the frequency interference product term corresponding to the distributed energy storage node numbered i and the tensor computation task numbered j at time t, which is a dimensionless value.

[0036] The frequency deviation squared term is obtained by subtracting the expected pulse angular frequency corresponding to the tensor calculation task at the same dimensionless moment from the electrochemical intrinsic angular frequency corresponding to the distributed energy storage node and taking the square. The frequency deviation squared term characterizes the degree of difference between the two frequencies; the smaller the value, the closer the two frequencies are, and the higher the probability of resonance. The formula for calculating the frequency deviation squared term is: in, Let be the squared frequency deviation term between the distributed energy storage node numbered i and the tensor computation task numbered j at time t, which is a dimensionless value.

[0037] The first derivative of the natural logarithm of normalized absolute temperature with respect to dimensionless time is calculated, and its squared value yields the transient temperature damping term. This term characterizes the damping effect of dynamic temperature changes on the resonance effect; the faster the rate of temperature change, the larger the transient temperature damping term, and the stronger its suppression of the resonance effect. For discretely sampled time-series data, the first derivative of the natural logarithm of normalized absolute temperature with respect to dimensionless time is calculated using the central difference method, while the initial time is calculated using the forward difference method. The formula for calculating the transient temperature damping term is: in, The temperature transient damping term corresponding to the distributed energy storage node numbered i at time t is a dimensionless value; an offset of 1e-6 is added to the logarithmic operation to avoid the error of undefined logarithm.

[0038] The resonance penalty factor is obtained by dividing the frequency interference product term by the sum of the frequency deviation square term and the temperature transient damping term, and then performing a definite integral operation in the dimensionless time domain from time zero to the current time. The definite integral operation is implemented using the trapezoidal integral method. For discretely sampled time series data, the integral calculation formula is: in, It is the integrand; The dimensionless sampling period is set to 1. The formula for calculating the resonance penalty factor is: in, The resonance penalty factor for the distributed energy storage node numbered i at time t and the tensor computation task numbered j is a dimensionless value; an offset of 1e-6 is added to the denominator to avoid division by zero. The integration operation is implemented using a sliding integration window with a length of 1000 dimensionless time units. Integration is performed only on the sampled data within the window to avoid numerical overflow during long-cycle operation.

[0039] The accumulated discharge energy is obtained by multiplying the normalized instantaneous current and normalized instantaneous voltage, and then performing a definite integral over the dimensionless time domain from time zero to the current time. The accumulated discharge energy characterizes the total discharge energy of the distributed energy storage node from the initial time to the current time, reflecting the node's remaining energy supply capacity. In the discharge state, the instantaneous current is positive, and the accumulated discharge energy increases with time; in the charging state, the instantaneous current is negative, and the accumulated discharge energy decreases with time. The formula for calculating the accumulated discharge energy is: in, Let be the cumulative discharge energy corresponding to the distributed energy storage node numbered i at time t, which is a dimensionless value. The integration operation adopts the same trapezoidal integration method and sliding integration window as the resonance penalty factor to avoid numerical overflow.

[0040] The maximum value among the resonance penalty factors for the corresponding distributed energy storage nodes under all tensor computation tasks is selected as the resonance penalty extreme value. The resonance penalty extreme value characterizes the maximum resonance risk of the distributed energy storage node under all currently assigned tensor computation tasks, and is used to constrain the node's computing power limit, preventing the node from undertaking computation tasks with high resonance risk. The formula for calculating the resonance penalty extreme value is: Where M is the total number of tensor computation tasks to be assigned; Let be the extremum of the resonance penalty corresponding to the distributed energy storage node numbered i at time t, which is a dimensionless value. When the calculated... When it is less than 1e-6, take This avoids errors such as division by zero in subsequent calculations.

[0041] The three-dimensional Laplace operator for the normalized absolute temperature in a three-dimensional temperature field is calculated, and the absolute value is used as the exponent of the natural constant for exponential operations to obtain the spatial heat dissipation decay term. The three-dimensional Laplace operator characterizes the spatial curvature of the absolute temperature distribution in three-dimensional space, reflecting the spatial diffusion characteristics of the temperature field. A larger value indicates a higher degree of non-uniformity in temperature distribution and a greater risk of localized heat accumulation. The spatial heat dissipation decay term is used to attenuate and correct the computing power capacity of nodes. A higher risk of localized heat accumulation corresponds to a larger value of the spatial heat dissipation decay term, resulting in a lower upper limit for the node's computing power capacity. The formula for calculating the three-dimensional Laplace operator is: in, Let be the three-dimensional Laplace operator corresponding to the distributed energy storage node numbered i at time t, and be a dimensionless value. For a discrete three-dimensional spatial grid, the second spatial derivative is calculated using the central difference method. The formula for calculating the spatial heat dissipation decay term is: in, Let be the spatial heat dissipation decay term corresponding to the distributed energy storage node numbered i at time t, which is a dimensionless value.

[0042] Dividing the cumulative discharge energy by the product of the resonance penalty extreme value and the spatial heat dissipation attenuation term, the computing power limit of the corresponding distributed energy storage node is derived. The calculation formula is as follows: in, Let t be the computing power limit of distributed energy storage node i at time t. It is a dimensionless value that represents the maximum computing power load that the node can safely handle. The larger the value, the higher the computing power load that the node can handle.

[0043] The variance of the resonance penalty factor, which comprises all distributed energy storage nodes and all tensor computation tasks, is calculated and divided by the expected value of the resonance penalty factor. The negative of this variance is then used as the exponent of the natural constant for exponential operations to obtain the dynamic inertia weight. The dynamic inertia weight is used in the iterative process of the particle swarm optimization algorithm to control the degree to which particles inherit the velocity of previous iterations. The higher the dispersion of the resonance penalty factor, the smaller the value of the dynamic inertia weight, the faster the particle's iteration velocity decays, and the stronger the algorithm's local optimization ability. Conversely, the lower the dispersion of the resonance penalty factor, the larger the value of the dynamic inertia weight, the stronger the particle's ability to maintain its iteration velocity, and the stronger the algorithm's global optimization ability. The formula for calculating the dynamic inertia weight is: in, The variance solver is used to calculate the variance of all resonance penalty factor samples; To find the desired solution operator, calculate the arithmetic mean of all samples of the resonance penalty factor; The dynamic inertia weight at time t is a dimensionless value ranging from 0 to 1; an offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0044] The dynamic cognitive learning factor is obtained by dividing the current distributed energy storage node's computing power boundary by the sum of the computing power boundaries of all distributed energy storage nodes. The dynamic cognitive learning factor controls the step size weight of a particle's learning towards its historical best position. The higher the proportion of the current node's computing power boundary, the larger the value of the dynamic cognitive learning factor, and the more inclined the particle is to iteratively update based on its own optimization experience. The formula for calculating the dynamic cognitive learning factor is: Where N is the total number of distributed energy storage nodes; Let i be the dynamic cognitive learning factor corresponding to the distributed energy storage node numbered i at time t. It is a dimensionless value with a range of 0 to 1. An offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0045] The dynamic social learning factor is obtained by subtracting the current distributed energy storage node's computing power boundary from the sum of the computing power boundaries of all distributed energy storage nodes, and then dividing the difference by the sum of the computing power boundaries of all distributed energy storage nodes. The dynamic social learning factor controls the step size weight of particles learning towards the global optimal position of the swarm. The lower the proportion of the current node's computing power boundary, the larger the value of the dynamic social learning factor, and the more the particles tend to iteratively update based on the swarm's optimization experience. The formula for calculating the dynamic social learning factor is: in, Let i be the dynamic social learning factor corresponding to the distributed energy storage node numbered i at time t. It is a dimensionless value with a range of 0 to 1. An offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0046] The third derivative of the normalized instantaneous voltage with respect to dimensionless time is calculated, and the sine value is taken and its absolute value is used as the first higher-order perturbation component. The third derivative of the normalized instantaneous current with respect to dimensionless time is calculated, and the cosine value is taken and its absolute value is used as the second higher-order perturbation component. The first and second higher-order perturbation components together constitute the physical perturbation quantity. The physical perturbation quantity is used to introduce random perturbations based on physical signals into the iterative process of the particle swarm optimization algorithm, preventing the algorithm from getting trapped in local optima. For discretely sampled time-series data, the third derivative of the normalized instantaneous voltage with respect to dimensionless time is approximated using the central difference method, and the calculation formula is as follows: For the initial three sampling times, the third derivative is calculated using the forward difference method. The third derivative of the normalized instantaneous current with respect to dimensionless time is calculated using the same difference method. The formula for calculating the physical disturbance is: in, This is the first high-order perturbation component. The second higher-order perturbation component consists of dimensionless values ​​ranging from 0 to 1; multiplied by the dimensionless sampling period. Ensure that the inputs for trigonometric functions are dimensionless quantities.

[0047] The particles in the particle swarm optimization algorithm are encoded using real numbers. Each particle corresponds to a set of distributed energy storage nodes and tensor computing tasks. The position vector of the particle is a one-dimensional vector after the two-dimensional allocation matrix is ​​expanded by rows. The dimension of the allocation matrix is ​​N×M, where N is the total number of distributed energy storage nodes and M is the total number of tensor computing tasks. The element in the i-th row and j-th column of the allocation matrix represents the computing power ratio of the tensor computing task numbered j to the distributed energy storage node numbered i. The allocation matrix satisfies the constraint that the sum of all elements in each column is 1, ensuring that the computing power requirement of each tensor computing task is fully allocated.

[0048] The basic operating parameters of the particle swarm optimization algorithm are set as follows: the particle population size is set to an integer between 50 and 200, the maximum number of iterations is set to an integer between 100 and 500, the convergence threshold is set to 1e-6, the particle position range is 0 to 1, and the particle evolution velocity range is -0.5 to 0.5. The initial positions of the particles are generated using a random uniform distribution, and after generation, each column is normalized to ensure that the sum of the elements in each column is 1; the initial evolution velocity of the particles is generated using a random uniform distribution, with a value range of -0.1 to 0.1.

[0049] The fitness function of the particle swarm optimization algorithm aims to minimize the sum of the resonance penalty factors corresponding to all distributed energy storage nodes and tensor computation tasks across the entire network. The formula for calculating the fitness function is as follows: Where X is the assignment matrix corresponding to the particle; To assign the element in the i-th row and j-th column of the matrix; The fitness value corresponds to the particle. The smaller the fitness value, the lower the resonance risk of the allocation scheme and the better the scheme.

[0050] This algorithm obtains the evolution velocity of the previous generation of particles, the individual difference distance between the current particle position and its best position, and the global difference distance between the current particle position and the global best position in the particle swarm optimization (PSO) algorithm. The individual best position is the position corresponding to the minimum fitness value achieved by a single particle in the iteration history; each particle maintains an independent individual best position, which is updated after each iteration. The global best position is the position corresponding to the minimum fitness value achieved by the entire particle swarm in the iteration history; the entire particle swarm maintains a unique global best position, which is updated after each iteration. The individual difference distance is the difference between the current particle position and its best position, and the global difference distance is the difference between the current particle position and the global best position.

[0051] Multiplying the evolution speed by the dynamic inertia weight, adding the product of the dynamic cognitive learning factor, the first higher-order perturbation component, and the individual difference distance, and then adding the product of the dynamic social learning factor, the second higher-order perturbation component, and the global difference distance, yields the updated evolution speed. The calculation formula is as follows: in, This represents the update evolution rate of particle k in the (n+1)th generation iteration; Let be the evolution velocity of the particle numbered k in the nth iteration; This represents the current position of the particle numbered k in the nth iteration. Let k be the optimal position for the individual particle numbered k. This represents the global optimal position of the particle swarm. When the calculated update evolution rate exceeds the range, boundary limiting is applied, setting the excess portion as the corresponding boundary value.

[0052] The particle position is updated based on the evolution velocity, and the position update formula is: in, This is the updated position of particle k in the (n+1)th iteration. If the updated particle position exceeds the range of values, boundary limiting is applied, setting the excess portion to the corresponding boundary value. After the position update, each column of the particle's allocation matrix is ​​normalized to ensure that the sum of all elements in each column is 1, satisfying the constraints of computational power allocation. The normalization formula is: An offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0053] After each iteration, the fitness values ​​of all particles are calculated, and the individual optimal position of each particle and the global optimal position of the particle swarm are updated. The update rule for the individual optimal position is: if the fitness value of the current particle is less than the fitness value corresponding to the individual optimal position, then the current particle position is updated to the new individual optimal position; the update rule for the global optimal position is: if the fitness value of the current particle is less than the fitness value corresponding to the global optimal position, then the current particle position is updated to the new global optimal position.

[0054] The iteration termination condition is set as follows: the iteration count reaches the preset maximum iteration count, or the fitness value corresponding to the global optimal position does not decrease for 20 consecutive iterations. The iteration terminates when either condition is met. After the iteration terminates, the allocation matrix corresponding to the global optimal position is taken as the optimal allocation matrix. Each element in the optimal allocation matrix corresponds to the computing power proportion of the tensor computation task numbered j allocated to the distributed energy storage node numbered i.

[0055] For the optimal allocation matrix, corresponding to a specific tensor computation task and a specific distributed energy storage node allocation combination, the partial derivative of the resonance penalty factor with respect to the expected pulse angular frequency is calculated as the resonance polarization gradient term. The resonance polarization gradient term characterizes the rate of change of the resonance penalty factor with the expected pulse angular frequency, reflecting the degree of influence of adjustments to the expected pulse angular frequency on resonance risk. The larger the value, the more significant the impact of small changes in the expected pulse angular frequency on resonance risk. For discrete sampled data, the partial derivative is approximated using the central difference method, and the calculation formula is: in, The number of a specific distributed energy storage node assigned to tensor computation task number j in the optimal allocation matrix; The step size for calculating the partial derivative is 1e-3; For the resonant polarization gradient term, is a dimensionless value.

[0056] The L2 norm of the set of partial derivatives of the specific distributed energy storage node under all tensor computation tasks is calculated, and then added to the absolute value of the first derivative of the corresponding normalized instantaneous voltage with respect to dimensionless time to form the dynamic voltage damping term. This dynamic voltage damping term is used to normalize the resonant polarization gradient term, eliminating gradient magnitude differences between different nodes and tasks, and simultaneously introducing a damping effect from dynamic voltage changes to prevent excessive fluctuations in the calculation results of the execution time offset. The formula for calculating the dynamic voltage damping term is: in, For the number The second norm of the set of partial derivatives of the distributed energy storage node under all tensor computation tasks is a dimensionless value. For the number The absolute value of the first derivative of the normalized instantaneous voltage of the distributed energy storage node with respect to dimensionless time is a dimensionless value. This is the dynamic voltage damping term, a dimensionless value. When the calculated... When it is less than 1e-6, take This avoids errors such as division by zero in subsequent calculations.

[0057] The execution time offset is obtained by dividing the resonant polarization gradient term by the dynamic voltage damping term and then calculating the arcsine value of the quotient. If the absolute value of the quotient is greater than 1, a limiting process is applied, restricting the quotient to the range of -1 to 1 to ensure that the arcsine function has a valid definition. The formula for calculating the execution time offset is: in, This is a limiting function that restricts the input value to the range of -1 to 1. Let j be the execution time offset corresponding to the tensor calculation task, which is a dimensionless radian value with a range of 1 / 2π. .

[0058] The number of empty instruction cycles is calculated based on the execution time offset, with the following dimension conversion rule: the dimensionless radian offset is converted to the physical time width, and then to the number of empty instruction cycles for the hardware acceleration unit. The formula for calculating the number of empty instruction cycles is: in, The number of empty instruction cycles corresponding to the tensor computation task with number j is a positive integer; The initial sampling period of the system is in seconds; The original dynamic clock frequency of the hardware acceleration unit during the runtime of the tensor computation task numbered j, in Hertz; This is the rounding function.

[0059] When the execution time offset is positive, a corresponding number of null instructions are inserted at the beginning of the tensor computation task's instruction stream, shifting the task's execution time backward. When the execution time offset is negative, a corresponding number of null instructions are inserted at the end of the tensor computation task's instruction stream, shifting the task's execution time forward. The null instruction insertion operation is completed through the underlying driver module. The insertion position avoids the intervals between computation nodes with data dependencies in the directed acyclic graph of the tensor computation task, ensuring that the insertion of null instructions does not disrupt the computational logic and data dependencies of the tensor computation task. In multi-task concurrent scenarios, the null instruction insertion operation is executed in descending order of task priority. The time shift operation of high-priority tasks is completed first, and the null instruction insertion position of low-priority tasks avoids the execution window of high-priority tasks, preventing hardware resource preemption and execution window conflicts.

[0060] The absolute value of the partial derivative of the normalized instantaneous voltage with respect to the normalized instantaneous current is calculated, multiplied by the square of the normalized instantaneous current, divided by the normalized absolute temperature, and then integrated over the dimensionless time domain. This result is multiplied again by the normalized absolute temperature, summed over all distributed energy storage nodes, and half the square of the summation is taken to obtain the fundamental thermodynamic term. The partial derivative of the normalized instantaneous voltage with respect to the normalized instantaneous current characterizes the dynamic internal resistance of the battery cell, reflecting the sensitivity of voltage to current changes. It is obtained by fitting the cell's volt-ampere characteristic curve using the least squares method. For discrete sampled data, the central difference method is used for approximate calculation. The calculation formula is as follows: in, The step size for calculating partial derivatives is 1e-3. The formula for calculating the fundamental thermodynamic terms is: in, Let t be the thermodynamic fundamental term corresponding to time t. It is a dimensionless value that represents the normalized value of the total heat dissipation energy generated by all distributed energy storage nodes during operation. The integration operation is implemented using the trapezoidal integration method and the sliding integration window to avoid numerical overflow. An offset of 1e-6 is added to the denominator to avoid the error of dividing by zero.

[0061] The system optimization cost term is obtained by squaring the Frobenius norm of the optimal allocation matrix and then performing a definite integral over the dimensionless time domain. This system optimization cost term characterizes the computational cost incurred by the system during the optimization process of computing power allocation, reflecting the complexity and computational overhead of the allocation scheme. The formula for calculating the system optimization cost term is: in, This is the optimal allocation matrix at time τ; The Frobenius norm of the optimal allocation matrix; Let be the system optimization cost term at time t, which is a dimensionless value; the integration operation is implemented using the trapezoidal integration method and the sliding integration window.

[0062] Adding the fundamental thermodynamic terms to the system optimization cost terms, we construct the energy evolution function, calculated as follows: in, Let t be the energy evolution function at time t, which is a dimensionless value that represents the combined level of total energy dissipation and computational cost during system operation.

[0063] The absolute value of the first derivative of the energy evolution function with respect to dimensionless time is calculated and integraled definitely over the dimensionless time domain as the heat dissipation feedback driving term. The heat dissipation feedback driving term characterizes the cumulative rate of change of system heat dissipation, reflects the thermal dynamic characteristics of the system during operation, and is used to drive the adaptive update of the sampling reference frequency. The formula for calculating the heat dissipation feedback driving term is: in, Let t be the heat dissipation feedback driving term corresponding to time t, which is a dimensionless value; the integration operation is implemented using the trapezoidal integration method and the sliding integration window.

[0064] The state-aware normalization term is obtained by multiplying the highest normalized absolute temperature of all distributed energy storage nodes by the lowest normalized instantaneous voltage of all distributed energy storage nodes. This state-aware normalization term is used to normalize the heat dissipation feedback driving term. Combined with the system's extreme operating states, it enables adaptive adjustment of the sampling reference frequency. The higher the system's highest temperature and the lower the lowest voltage, the larger the value of the state-aware normalization term, the smaller the update amplitude of the sampling reference frequency, and the more stable the system operation. The formula for calculating the state-aware normalization term is: in, Let be the state-aware normalization term corresponding to time t, which is a dimensionless value. When the calculated... When it is less than 1e-6, take This avoids errors such as division by zero in subsequent calculations.

[0065] Dividing the heat dissipation feedback driving term by the state-aware normalization term updates the sampling reference frequency for the next sampling period. The calculation formula is as follows: in, This is the sampling reference frequency for the next sampling period, in Hertz; The system reference sampling frequency is set to 1 kHz and is a fixed constant. Upper and lower limits are set for updating the sampling reference frequency: an upper limit of 1 MHz and a lower limit of 100 Hz. If the calculated sampling reference frequency exceeds these limits, boundary limiting is applied to confine it within the specified range. The maximum adjustment ratio for a single sampling reference frequency is set to 20%, meaning the ratio of the updated sampling reference frequency to the current sampling reference frequency must not exceed 1.2 and must not be lower than 0.8, to avoid closed-loop system oscillations caused by large changes in the sampling frequency.

[0066] After the sampling reference frequency is updated, the new sampling period parameters are sent to the acquisition modules, analog-to-digital converters and hardware acceleration units of all distributed energy storage nodes through the master clock source. All modules synchronously update the sampling period and trigger pulse interval to ensure the sampling synchronization of the entire system.

[0067] The distributed multi-node collaborative scheduling adopts a centralized architecture. The system's computing and processing unit acts as the central scheduling node, responsible for global data acquisition, parameter calculation, computing power allocation, and scheduling command issuance. All distributed energy storage nodes and hardware acceleration units act as slave nodes, receiving commands from the central scheduling node and executing corresponding operations. The redundancy switching mechanism for node failures is as follows: when a distributed energy storage node goes offline, loses acquired data, or its computing power capacity reaches zero, the central scheduling node reallocates the computing power tasks corresponding to that node to other normally operating distributed energy storage nodes, re-executes the particle swarm optimization algorithm to update the optimal allocation matrix, and ensures the continuous and stable operation of the system.

[0068] The abnormal operating condition handling mechanism includes: when the acquisition module loses data, linear interpolation is performed to complete the data using the valid data from the previous moment; when the hardware acceleration unit fails, the corresponding tensor calculation task is migrated to another normally operating hardware acceleration unit; when the resonance penalty factor exceeds the preset threshold, the computing power undertaking ratio of the corresponding node is immediately reduced, triggering the task redistribution process to avoid cell degradation caused by resonance.

[0069] The system initialization process is performed according to the following steps to ensure that the system can stably enter a closed-loop operation state after startup.

[0070] After the system powers on, it first executes a hardware self-test procedure, sequentially performing self-tests on the analog-to-digital converter, the acquisition modules of each distributed energy storage node, the hardware acceleration unit, and the underlying driver module. This confirms that the communication links, sampling functions, and command issuance functions of all hardware modules are normal. During the hardware self-test, if any module fails to self-test, the system will record the faulty module's number and fault type, and disable the corresponding functions of that faulty module, using only the modules that passed the self-test to complete subsequent initialization and operation procedures.

[0071] After the hardware self-test is completed, the parameter initialization process is executed. The rated voltage, rated discharge current, and spatial coordinate parameters of all distributed energy storage nodes are read. The rated maximum main frequency and the number of arithmetic logic units of the hardware acceleration unit are read. The system reference sampling frequency, the basic operating parameters of the particle swarm algorithm, the sliding integral window length, the upper and lower limits of the sampling frequency update, and the single maximum adjustment ratio parameters are set. All parameters are stored in the system's non-volatile storage unit, supporting online modification and updates during operation.

[0072] After parameter initialization, the initial data acquisition and benchmark calibration process is executed. The system continuously acquires instantaneous voltage, instantaneous current, and absolute temperature data of all distributed energy storage nodes for 100 cycles at a benchmark sampling frequency. After removing outliers from the acquired data, the initial electrochemical intrinsic angular frequency benchmark value of each distributed energy storage node is calculated. Simultaneously, the instantaneous utilization benchmark value of the arithmetic logic unit under no-load conditions of the hardware acceleration unit is acquired, completing the parameter calibration of the system's initial state. Outlier removal adopts the Laida criterion, identifying acquired data exceeding three standard deviations of the average value as outliers, and replacing outliers with the linear interpolation result of adjacent valid data.

[0073] After initial calibration, the system enters a waiting-for-scheduling state, awaiting the delivery of tensor computation tasks. Upon receiving a new tensor computation task, the system first parses the directed acyclic graph of the task, extracts the tensor parameters of the core computation nodes, calculates the expected pulse angular frequency corresponding to the task, and then executes a closed-loop process including resonance penalty factor calculation, computational power boundary convergence, particle swarm optimization to solve for the optimal allocation matrix, task time shifting, heat dissipation feedback, and sampling frequency update.

[0074] The complete timing sequence of the system's closed-loop operation is executed according to a fixed scheduling cycle. The operation steps within each scheduling cycle are executed in the following order to ensure that the timing dependencies of all operations are correct and there are no logical conflicts.

[0075] At the start of the scheduling cycle, the system sends a synchronous trigger pulse through the master clock source. The acquisition modules and analog-to-digital converters of all distributed energy storage nodes synchronously complete the acquisition of instantaneous voltage, instantaneous current and absolute temperature data for the current cycle, complete the normalization processing and outlier removal of the acquired data, and upload the processed valid data to the system's computing unit.

[0076] After data acquisition is completed, the system constructs a three-dimensional spatial temperature field for the current cycle based on the acquired valid data, calculates the electrochemical intrinsic angular frequency corresponding to each distributed energy storage node, and simultaneously analyzes all tensor calculation tasks to be executed in the current cycle, calculating the expected pulse angular frequency corresponding to each task.

[0077] After the parameters are calculated, the system calculates the resonance penalty factor corresponding to each distributed energy storage node and each tensor computation task combination, converges to obtain the computing power bearing boundary of each distributed energy storage node in the current cycle, and updates the dynamic inertia weight, dynamic cognitive learning factor and dynamic social learning factor of the particle swarm algorithm.

[0078] After the parameters are updated, the system executes the iterative optimization process of the particle swarm optimization algorithm to solve for the optimal allocation matrix of the current cycle and determine the ratio of distributed energy storage nodes and computing power allocation for each tensor computation task.

[0079] After the optimal allocation matrix is ​​solved, the system calculates the execution time offset for each tensor computation task, determines the number of empty instruction cycles and the insertion position, and inserts empty instructions into the instruction stream of the corresponding tensor computation task through the underlying driver module to complete the translation operation of task execution time.

[0080] After the time shift operation is completed, the system calculates the energy evolution function and heat dissipation feedback driving term based on the allocation matrix and collected data of the current period, updates the sampling reference frequency of the next sampling period, and sends the updated sampling period parameters to all hardware modules through the master clock source, thus completing all operations of the current scheduling period.

[0081] After the operation of the current scheduling cycle is completed, the system enters the waiting state of the next scheduling cycle. At the beginning of the next scheduling cycle, the above complete process is repeated to achieve continuous closed-loop operation of the system.

[0082] The priority scheduling rules for tensor computation tasks are executed as follows to ensure the stable execution of high-priority tasks and low resonance risk. Each tensor computation task is assigned a corresponding priority level, which is divided into four levels: real-time priority, high priority, normal priority, and low priority. The priority level is specified synchronously when the task is issued, with real-time priority being the highest priority and low priority being the lowest priority.

[0083] For real-time priority tensor computation tasks, the system prioritizes ensuring their computing power requirements. During the optimal allocation matrix solution process, the distributed energy storage node with the smallest resonance penalty factor is prioritized for real-time priority tasks to ensure the execution stability of the task and the lowest resonance risk. The time translation operation of real-time priority tasks is only allowed to be executed within the task's idle waiting window and must not change the total execution time and deadline of the task.

[0084] For high-priority tensor computation tasks, the system ensures the computing power requirements during allocation while prioritizing distributed energy storage nodes with resonance penalty factors below a preset threshold. The maximum number of empty instruction cycles allowed for time shift operations of high-priority tasks does not exceed 5% of the total execution cycle of the task, thus avoiding excessive delays in task execution.

[0085] For tensor computation tasks of normal and low priority, after meeting the computing power requirements of high priority and real-time priority tasks, the system completes the computing power allocation based on the goal of minimizing the total sum of network resonance penalties. The maximum number of empty instruction cycles allowed for time translation operations of normal priority tasks does not exceed 20% of the total execution cycle of the task, while there is no maximum cycle limit for time translation operations of low priority tasks. The task execution time can be flexibly adjusted based on resonance avoidance requirements.

[0086] When a new tensor computation task is issued, if the remaining computing power of all distributed energy storage nodes cannot meet the computing power requirements of the new task, the system will suspend or reduce the computing power allocation of low-priority tasks in order of priority from low to high, and give priority to ensuring the computing power requirements of high-priority and real-time priority tasks. The suspended low-priority tasks will be re-executed after the computing power resources are released.

[0087] The online calibration process for battery cell parameters is executed according to a fixed calibration cycle, which is set to 1000 scheduling cycles to ensure that the calculation accuracy of the electrochemical intrinsic angular frequency matches the actual characteristics of the battery cell. During the online calibration process, the system collects complete charge-discharge cycle data of all distributed energy storage nodes within the current calibration cycle. Based on the collected volt-ampere characteristic data, the dynamic internal resistance characteristic curve of the battery cell is refitted, and the calculation parameters of the partial derivative of the normalized instantaneous voltage with respect to the normalized instantaneous current are updated. At the same time, based on the current health status of the battery cell, the calibration coefficients of the rated voltage and rated discharge current of the battery cell are updated to ensure the accuracy of subsequent parameter calculations.

[0088] The health status of a battery cell is assessed based on cumulative discharge energy, cumulative operating time, and the rate of change of the cell's internal resistance. The formula for calculating the rate of change of the cell's internal resistance is as follows: in, The current internal resistance value of the distributed energy storage node numbered i at time t is obtained by fitting the volt-ampere characteristic. Let i be the initial internal resistance value of the distributed energy storage node numbered i. The cell internal resistance change rate is a dimensionless value. When the cell internal resistance change rate exceeds 20%, the system will reduce the maximum computing power limit of that node to prevent the cell from aging rapidly under high load.

[0089] The convergence guarantee measures for the Particle Swarm Optimization (PSO) algorithm are implemented as follows to avoid the algorithm getting trapped in local optima and ensure that the optimal allocation matrix obtained is the global optimum. During the algorithm iteration process, if the fitness value corresponding to the global optimum position does not decrease for 10 consecutive iterations, the system will reinitialize 30% of the particles with poor fitness values ​​in the population. The positions of the reinitialized particles are generated using a random uniform distribution, and the dynamic inertia weight is temporarily adjusted to the initial value of 0.9 to expand the global search range of the algorithm and avoid getting trapped in local optima.

[0090] During the algorithm iteration process, the dynamic inertia weight adopts an adaptive decay strategy. As the number of iterations increases, a linear decay factor is additionally added to the original physical drive of the dynamic inertia weight. The formula for calculating the linear decay factor is as follows: in, This represents the current iteration number; This represents the maximum number of iterations. The maximum inertia weight is set to 0.9. The minimum inertia weight is set to 0.4. The linear decay factor is used. The final dynamic inertia weight is the product of the physics-driven dynamic inertia weight and the linear decay factor, ensuring that the algorithm has strong global search capability in the early stage of iteration and strong local optimization capability in the later stage of iteration, thereby improving the convergence speed and solution accuracy of the algorithm.

[0091] The system's closed-loop stability assurance measures are implemented as follows to avoid positive feedback divergence during sampling frequency updates and ensure the stability of the system's closed-loop operation. After each update of the sampling reference frequency, the system will record the changing trends of the system's highest temperature, lowest voltage, and the sum of the resonance penalty factors over 10 consecutive scheduling cycles. If the system's highest temperature continues to rise or the sum of the resonance penalty factors continues to increase over 3 consecutive scheduling cycles, the system will immediately suspend the adaptive update of the sampling reference frequency, revert to the system's reference sampling frequency, and trigger a task reallocation process to reduce the computing load on nodes with high resonance risk. The adaptive update function of the sampling reference frequency will be restarted only after the system's operating state has stabilized.

[0092] During system operation, all collected data, calculated parameters, allocation matrices, and scheduling instructions are stored according to fixed timestamps, with the storage period consistent with the scheduling period. The stored data can be used for subsequent system operation status analysis, fault tracing, and parameter optimization. The stored data is managed using a cyclic overwrite method. When the storage space reaches a preset threshold, the system will automatically delete the oldest stored historical data, retaining only the most recent operational data and critical fault event data, ensuring the stable availability of system storage resources.

[0093] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An AI computing power optimization and scheduling system for new energy storage, applied to a system comprising distributed energy storage nodes and a hardware acceleration unit with a single-instruction multi-threaded architecture and an arithmetic logic unit, wherein the system includes an analog-to-digital converter, characterized in that, include: Real-time acquisition of instantaneous voltage, instantaneous current and absolute temperature of each node is used to construct a three-dimensional temperature field, and the electrochemical intrinsic angular frequency is calculated in reverse. The directed acyclic graph of multiple tensor computation tasks sent to the hardware acceleration unit is parsed, and the tensor dimension size and dynamic main frequency are extracted and mapped to the expected pulse angular frequency. The resonance penalty factor is constructed by performing time-domain integration between the expected pulse angular frequency and the electrochemical intrinsic angular frequency. Based on the resonance penalty factor and the three-dimensional spatial temperature field, the computing power bearing boundary of each node is converged. The higher-order derivatives of the instantaneous voltage and instantaneous current are extracted as physical perturbations. The optimal allocation matrix is ​​obtained by using the computing power to drive the particle swarm optimization algorithm to find the optimal allocation matrix. Calculate the execution time offset of each task in the optimal allocation matrix, and dynamically insert empty instructions into the underlying drive flow to perform time shifting of the tensor calculation task; The heat dissipation caused by the translation task is encapsulated in the energy evolution function, and its transient derivative is used to update the sampling reference frequency of the analog-to-digital converter in reverse.

2. The AI ​​computing power optimization and scheduling system for new energy storage according to claim 1, which collects the instantaneous voltage, instantaneous current and absolute temperature of each of the distributed energy storage nodes in real time, and calculates the electrochemical intrinsic angular frequency in reverse, includes: Multiply the first derivative of the instantaneous voltage with respect to time by the instantaneous current, subtract the product of the first derivatives of the instantaneous voltage and the instantaneous current with respect to time, and take the absolute value of the difference as the amplitude of the phase angle change rate. The square of the instantaneous voltage is added to the square of the instantaneous current to obtain the sum of the squares of the electrical signal amplitude; Divide the amplitude of the phase angle change rate by the sum of the squares of the electrical signal amplitudes to obtain the reference angular frequency component; Divide the second derivative of the absolute temperature with respect to time by the absolute temperature, and use the resulting quotient as the exponent of the natural constant for exponential operation to obtain the thermal stimulation modulation term. Multiplying the reference angular frequency component by the thermally excited modulation term yields the electrochemical intrinsic angular frequency corresponding to the distributed energy storage node.

3. The AI ​​computing power optimization and scheduling system for new energy storage according to claim 1, comprising parsing the directed acyclic graph of multiple tensor computing tasks sent to the hardware acceleration unit, extracting the tensor dimension size and the dynamic main frequency of the hardware acceleration unit, and mapping them across dimensions to the expected pulse angular frequency, including: Extract the total number of tensor dimensions and the size of the tensor dimensions in the tensor computation task; Summing the sizes of all the tensor dimensions yields the sum of the tensor dimensions and values. Multiply all the tensor dimensions together, and take the arithmetic root of the total number of dimensions of the product to obtain the geometric mean of the tensor dimensions. Extract the instantaneous utilization rate of the arithmetic logic unit, calculate the first derivative of the natural logarithm of the instantaneous utilization rate with respect to time and take its absolute value, then add one to obtain the dynamic load rate derivative term; The expected pulse angular frequency corresponding to the tensor computation task is obtained by continuously multiplying the dynamic main frequency, the tensor dimension sum and value, and the dynamic load rate derivative term, and then dividing by the geometric mean of the tensor dimension.

4. The AI ​​computing power optimization and scheduling system for new energy storage according to claim 1, wherein the expected pulse angular frequency and the electrochemical intrinsic angular frequency are integrated in the time domain to construct a resonance penalty factor, comprising: Multiplying the expected pulse angular frequency by the electrochemical intrinsic angular frequency yields the frequency interference product term; Subtracting the expected pulse angular frequency from the electrochemical intrinsic angular frequency and squaring the result yields the squared frequency deviation term. Calculate the first derivative of the natural logarithm of the absolute temperature with respect to time, and square it to obtain the transient temperature damping term; The resonance penalty factor is obtained by dividing the frequency interference product term by the sum of the frequency deviation square term and the temperature transient damping term, and then performing a definite integral operation in the time domain from time zero to the current time.

5. The AI ​​computing power optimization and scheduling system for new energy storage according to claim 1, based on the resonance penalty factor and the three-dimensional spatial temperature field, converges to obtain the computing power bearing boundary of each node, including: Multiply the instantaneous current by the instantaneous voltage and perform a definite integral operation in the time domain from time zero to the current time to obtain the cumulative discharge energy; The maximum value among the resonance penalty factors corresponding to the distributed energy storage node under all tensor computation tasks is selected as the resonance penalty extreme value; Calculate the three-dimensional Laplace operator of the absolute temperature in the three-dimensional temperature field, and take the absolute value as the exponent of the natural constant to perform exponential operation to obtain the spatial heat dissipation decay term; By dividing the cumulative discharge energy by the product of the resonance penalty extreme value and the spatial heat dissipation attenuation term, the computing power bearing boundary corresponding to the distributed energy storage node is derived.

6. The AI ​​computing power optimization scheduling system for new energy storage according to claim 1, comprising extracting the higher-order derivatives of the instantaneous voltage and instantaneous current as physical perturbations, and using the computing power boundary to drive the optimization iteration of particles in the particle swarm optimization algorithm to solve for the optimal allocation matrix, including: Calculate the variance of the resonance penalty factor composed of all distributed energy storage nodes and all tensor computation tasks, divide it by the expected value of the resonance penalty factor, take the negative number and use it as the exponent of the natural constant for exponential operation to obtain the dynamic inertia weight. Divide the computing power boundary of the current distributed energy storage node by the sum of the computing power boundaries of all distributed energy storage nodes to obtain the dynamic cognitive learning factor. The difference between the sum of the computing power acceptance boundaries and the current computing power acceptance boundaries of the distributed energy storage nodes is divided by the sum of the computing power acceptance boundaries to obtain the dynamic social learning factor. Obtaining the physical disturbance quantity specifically includes: calculating the third derivative of the instantaneous voltage with respect to time, taking the sine value and calculating the absolute value as the first higher-order disturbance component; and calculating the third derivative of the instantaneous current with respect to time, taking the cosine value and calculating the absolute value as the second higher-order disturbance component. Obtain the evolution velocity of the previous generation of particles, the individual difference distance between the current particle position and the individual optimal position, and the global difference distance between the current particle position and the global optimal position in the particle swarm algorithm. Multiply the evolution speed by the dynamic inertia weight, add the product of the dynamic cognitive learning factor, the first higher-order perturbation component and the individual difference distance, and add the product of the dynamic social learning factor, the second higher-order perturbation component and the global difference distance to obtain the updated evolution speed. The optimal allocation matrix that minimizes the total network penalty is obtained by iteratively solving using the aforementioned update evolution speed.

7. The AI ​​computing power optimization scheduling system for new energy storage according to claim 1, comprising calculating the execution time offset of each task in the optimal allocation matrix, and dynamically inserting empty instructions into the underlying drive flow to perform time shifting of the tensor calculation task, including: For the allocation combination of a specific tensor computation task and a specific distributed energy storage node in the optimal allocation matrix, calculate the partial derivative of the resonance penalty factor with respect to the expected pulse angular frequency, and use it as the resonance polarization gradient term; Calculate the L2 norm of the set of partial derivatives of the specific distributed energy storage node under all tensor computation tasks, and add the absolute value of the first derivative of the instantaneous voltage with respect to time to form the dynamic voltage damping term; Divide the resonant polarization gradient term by the dynamic voltage damping term, and calculate the arcsine value of the quotient to obtain the execution time offset. The empty instruction cycle corresponding to the time width is calculated based on the execution time offset, and the empty instruction corresponding to the empty instruction cycle is dynamically inserted into the underlying driver stream to complete the time shift.

8. The AI ​​computing power optimization and scheduling system for new energy storage according to claim 1, wherein the heat dissipation caused by the shifted task is encapsulated in the energy evolution function, and its transient derivative is used to update the sampling reference frequency of the analog-to-digital converter in reverse, comprising: The absolute value of the partial derivative of the instantaneous voltage with respect to the instantaneous current is calculated, multiplied by the square of the instantaneous current, divided by the absolute temperature, and then integrated in the time domain. This result is multiplied by the absolute temperature again, summed over all distributed energy storage nodes, and half the square is taken to obtain the fundamental thermodynamic term. The Frobenius norm of the optimal allocation matrix is ​​squared and then integrated in the time domain to obtain the system optimization cost term. The energy evolution function is constructed by adding the fundamental thermodynamic terms to the system optimization cost terms. Calculate the absolute value of the first derivative of the energy evolution function with respect to time, and integrate it in the time domain as a heat dissipation feedback driving term; Multiply the highest absolute temperature among all distributed energy storage nodes by the lowest instantaneous voltage among all distributed energy storage nodes to obtain the state-aware normalization term. Divide the heat dissipation feedback driving term by the state-aware normalization term to update the sampling reference frequency for the next sampling period.

Images