Electric vehicle hybrid energy storage system energy management method fused with Hemma particle swarm optimization LSTM

Abstract

The invention discloses an electric vehicle hybrid energy storage system energy management method fused with a Hemma particle swarm algorithm LSTM, and the method comprises the steps: constructing a multi-layer LSTM neural network controller to learn the characteristics of a driving condition, and minimizing the stress of a battery; constructing a fractional order integral derivative (FOID) controller, outputting a power correction component, and decomposing the component into a battery power reference component and a supercapacitor power reference component; according to the method, an improved He-Marx particle swarm algorithm is utilized, LSTM weight and FOID controller parameters are optimized in a combined mode through a multi-objective optimization function (including battery root-mean-square current and voltage errors), optimization training is carried out on the weight and bias of a neural network controller, and power distribution between a battery and a supercapacitor in the hybrid energy storage system of the electric vehicle is controlled in real time; an LSTM power distribution result is adaptively corrected through a dynamic weight mechanism, and bus voltage fluctuation is controlled within a range of + / -3%; and the final power reference value is synthesized through dynamic weighting, so that the peak current of the battery is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electric vehicle energy management, in particular to an electric vehicle hybrid energy storage system energy management method fusing a river horse particle swarm algorithm and an LSTM (Long Short-Term Memory), which is particularly suitable for a battery-super capacitor hybrid energy storage system and realizes multi-scale power distribution through the collaborative optimization of fractional order control and the LSTM. BACKGROUND

[0002] The development of electric vehicles (EVs) is crucial for reducing carbon emissions in the transportation sector. However, improving their economic efficiency and range requires addressing challenges such as high battery costs and limited lifespan. When dealing with large power transients during vehicle acceleration, braking, and other operating conditions, batteries are subjected to high currents and rapid charge-discharge cycles. Combining batteries with supercapacitors to form a hybrid energy storage system can effectively reduce battery stress and prolong its lifespan. This is an important strategy for improving the performance of electric vehicles.

[0003] The energy management system is responsible for intelligently distributing power between the battery and the supercapacitor. Offline optimization strategies can provide a design basis for online strategies such as rule-based or learning-based strategies. Long Short-Term Memory (LSTM) has shown potential in the field of energy management and is used to predict energy consumption or learn optimal control strategies. Existing PID controllers have limitations in suppressing voltage fluctuations, such as rigid frequency response and insufficient disturbance rejection capability.

[0004] The application introduces a fractional order integral derivative (FOID) controller, which realizes the continuous switching of PI / PD characteristics through parameters (lambda∈[0.1,0.9]) to improve transient response and robustness. Therefore, there is an urgent need to develop a hybrid energy storage system energy management method and system that can effectively reduce battery stress, prolong its lifespan, ensure stable DC bus voltage, and be easily optimized and deployed. SUMMARY

[0005] The application aims to overcome the shortcomings of the prior art and provide an electric vehicle hybrid energy storage system energy management method fusing a river horse particle swarm algorithm and an LSTM. The method combines LSTM and fractional order integral derivative (FOID) controllers, dynamically balances battery stress minimization and voltage stability targets through a dynamic weight mechanism eta(t), and optimizes FOID controller parameters through an improved fusion particle swarm river horse optimization algorithm to achieve optimal (or suboptimal) power distribution between the battery and the supercapacitor, significantly reducing battery root mean square current and peak current, reducing battery stress, and prolonging its lifespan.

[0006] Technical solution: The application discloses an electric vehicle hybrid energy storage system energy management method fusing a river horse particle swarm algorithm and an LSTM, which comprises the following steps:

[0007] Step 1: Obtain electric vehicle driving cycle data, including vehicle speed V and load power P. l ;

[0008] Step 2: Construct a multi-layer LSTM neural network controller, and process the historical state data (velocity V, load power P) l Supercapacitor voltage U c DC bus voltage error As input, the time-dependent features of driving conditions are learned, and the multi-layer LSTM neural network controller learns the mapping from the input state to power allocation, outputting the battery power reference component. and supercapacitor power reference component

[0009] Step 3: Construct a fractional-order integral derivative FOID controller. The FOID controller uses the DC bus reference voltage. With the measured voltage U DC error As input, the output power correction component P FOID , will P FOID Decomposed into battery power reference components and supercapacitor power reference component Used to correct the power distribution results of the multilayer LSTM neural network controller, and control the bus voltage fluctuation within ±3% of the reference value;

[0010] Step 4: Based on the optimization algorithm, a multi-objective joint optimization framework is adopted. The primary objective is to minimize the battery root mean square current, and the weights and biases of the multilayer LSTM neural network controller in Step 2 are optimized and trained. The secondary objective is to minimize the voltage error, and the FOID controller parameters (K) are optimized accordingly. i ,K d Optimization of λ); the objective function is:

[0011] Fitness = α·I bat,RMS +β·|E vdc | (1)

[0012] in, I bat (t) represents the battery current time series, T is the simulation time, α and β are the weighting coefficients of the primary and secondary objectives, α + β = 1, |E vdc | represents the Laplace transform of the voltage error, λ∈[0.1,0.9];

[0013] Step 5: Convert the power reference component obtained in Step 2 With the power reference component from step 3 Dynamic weighted summation yields the final battery power reference value. and supercapacitor power reference value And input to the battery and supercapacitor DC / DC converter.

[0014] Furthermore, in step 1, the load power P l The calculation method is as follows:

[0015] Multiple acceleration and constant speed deceleration intervals are set up: acceleration intervals of 1–15s and 48–55s, constant speed intervals of 15–20s and 55–60s, and deceleration intervals of 20–30s and 60–65s. The driving of a car is simulated, and the electrical load P required for driving at each time point is calculated. l The details are as follows:

[0016]

[0017] Among them, F m As the driving force, W tg F is the tangent component of gravity (assuming a slope of 0). air For air resistance, F roll For rolling resistance, P m The mechanical power at the wheel is ρ, where ρ is the air density and C is the density of air. d Let A be the drag coefficient. f Where η is the vehicle's frontal area, η is the electromechanical conversion power, V is the driving speed, and C is the vehicle's frontal area. rr This is the tire deformation dissipation coefficient.

[0018] Furthermore, step 4 utilizes an improved fusion particle swarm optimization algorithm, as detailed below:

[0019] Step 4.1: Initialize PSO algorithm parameters, including the number of particles / individuals n, the maximum number of iterations iter, the inertia weight u, the initial inertia u0, and the final inertia value u. f Set the catastrophic perturbation probability and the proportion of elite particles retained in the HO algorithm, with the upper and lower bounds of the particle position in each dimension being x. max x min ;

[0020] Step 4.2: The position of each particle / individual represents a set of weights and biases of the multilayer LSTM neural network controller with respect to FOID K. i ,K d ,λ; Hippo particle swarm fusion formula initialization of particle swarm:

[0021] x t =x min +r[(rand*x max -0.5)-x min (3)

[0022] Where r is a random number between 0 and 1, and rand is a random number;

[0023] The nonlinear decaying inertia weight is:

[0024] u = u0 - (u0-u f )*iter / n (4)

[0025] The formula for updating the position of each particle / individual is:

[0026] x iter = u*x0 + c1*r1*(x b -x0) + c2*r2*( x g - x0 ) (5)

[0027] Where, x b For the optimal value of the particle, x g The global optimal value is represented by c1 and c2, which are acceleration factors, and r1 and r2 are social cognition coefficients.

[0028] Step 4.3: Compare the fitness obtained from the electrical simulation with the fitness corresponding to the best position in the particle's history, and set the calculation benchmark. If the particle gets stuck, use the HO algorithm to perform a local search and perturb the non-elite particles. If the particle gets stuck multiple times, keep the elite particles and reset the rest.

[0029] In the HO algorithm, elite particles represent male hippos, and the growth mechanism for non-elite particles is improved by introducing a new convergence factor T.

[0030]

[0031] Where, x t+1 This indicates the position after the HO algorithm performs a local search, t is the current iteration number, iter is the maximum iteration number, A is a random number between 0 and 1, and B is a random number between 1 and 2. When T > 0.9, the naive hippo tends to perform a fine search (local development) near the optimal solution, thus improving the algorithm's accuracy.

[0032] Step 4.4: For each particle / individual position, decode into a set of weights and biases of a multilayer LSTM neural network controller, and the parameters K of FOID. i ,K d ,λ;

[0033] Step 4.5: Perform electrical simulation using a set of weights and biases of a multilayer LSTM neural network controller under acceleration / deceleration driving cycles or target driving cycles, and use the decoded FOID parameter K. i ,K d ,λ is used to obtain the Laplace transform of the voltage error in the electrical simulation;

[0034] Step 4.6: Calculate the objective function value as the fitness;

[0035] Step 4.7: Update the optimal position and global optimal position of the particle / individual.

[0036] Step 4.8: Repeat steps 4.2-4.7 until the maximum number of iterations is reached or convergence is achieved. Output a set of weights and biases of the multilayer LSTM neural network controller corresponding to the global optimal position, as well as the parameter K of FOID. i ,K d ,λ.

[0037] Furthermore, the multilayer LSTM neural network controller adopts a multilayer structure, including an input layer V, P... l U c and 50-cell LSTM layer, 40-cell fully connected layer and output layer, outputting battery power reference component. and supercapacitor power reference component

[0038] Furthermore, the multi-layer LSTM neural network controller learns the time-dependent features of driving conditions, specifically as follows:

[0039] LSTM solves the dependency problem through a gating mechanism. The core formula is as follows:

[0040] Input gate: i t =σ(W i [h t-1 ,x t ]+b i )

[0041] Forgotten Gate: f t =σ(W f [h t-1 ,x t ]+b f )

[0042] Candidate memories:

[0043] Memory update:

[0044] Output gate: o t =σ(W o [h t-1 ,x t ]+b o )

[0045] Hidden layer output: h t =o t tanh(C t )

[0046] h is the current input. t-1 The hidden state is the state at the previous time step, ⊙ represents element-wise multiplication, σ and tanh are the sigmoid activation functions, and W is the hidden state at the previous time step. i W f W c W o These are the weights for the input gate, forget gate, candidate memory, and output gate, respectively, b. i ,b f ,b c ,b o These are the input gate, forget gate, candidate memory, and output gate biases, respectively, n. h ,n x To hide the number of units and the dimension of the input vector, a state persistence mechanism is used to retain historical state information to solve the temporal dependency problem. The state update equation is as follows:

[0047] ctht=ft ct-1+it ct~=ot tanh(ct) (8)

[0048] Where: ft: forget gate output, it: input gate output, ct~: candidate cell state, trained through LSTM layers and fully connected layers, finally outputting the battery power reference component. and supercapacitor power reference component

[0049] Furthermore, the FOID controller in step 3 is specifically as follows:

[0050] Fractional operator s of FOID controller λ Discretization was achieved using a 5th-order Oustaloup filter, and integral operators H were constructed respectively. i (s), corresponding to s -λ Differential operator H d (s), corresponding to s 1-λ :

[0051]

[0052] H i (z)≈s -λ H d (z)≈s 1-λ

[0053] Among them, T z = 1ms is the sampling period, ω k ,ω' k For the recursive calculation of poles / zeros, K is the amplitude compensation gain, optimized through zero-pole matching, ω l ω is the low-frequency cutoff frequency. hThis is the high-frequency cutoff frequency;

[0054] Input DC bus voltage With the measured voltage U DC error

[0055] Output power correction component:

[0056]

[0057] P FOID Decomposed into battery power reference component and supercapacitor power reference component:

[0058]

[0059] Among them, K B and K C Used to adjust the dominance of the PID controller in power distribution between the battery and the supercapacitor, E vdc (s) is the Laplace transform of the voltage error, K i K is the integral gain, used for the integral term of the error, and is used to eliminate steady-state error. d The differential gain, used for the differential term of the error, is used for constant instantaneous fluctuations, L -1 λ is the inverse Laplace transform operator, and λ is the fractional calculus intermediate order, with values ​​ranging from [0.1, 0.9].

[0060] Furthermore, the output component of the FOID controller ( and ) respectively through the weight constant (K B ,K C After scaling, it participates in the dynamic weighted summation:

[0061]

[0062] Among them, dynamic weights The sensitivity coefficient k = 5.0, when |E vdc When |>2V, η(t)>0.88, FOID dominates voltage regulation.

[0063] Beneficial effects:

[0064] This invention uses a multi-layer LSTM neural network controller to learn and train the time-dependent characteristics of driving conditions, learning the mapping from input states to power allocation, and determining the battery power reference components and supercapacitor power reference components under the LSTM neural network. A fractional-order integral derivative FOID controller is used to determine the output power correction component, correcting the power allocation results of the multi-layer LSTM neural network controller, and determining the battery power reference components and supercapacitor power reference components corresponding to the FOID controller. Finally, through a dynamic weighting mechanism η(t), the optimal (or suboptimal) power allocation between the battery and supercapacitor is achieved by adaptively balancing the battery stress minimization and voltage stability objectives. This significantly reduces the battery's root-mean-square current and peak current, alleviates battery stress, and extends its service life. Attached Figure Description

[0065] Figure 1 For the two acceleration events of the electric vehicle;

[0066] Figure 2 Composed of DC bus current;

[0067] Figure 3 This is the DC bus power configuration;

[0068] Figure 4 Generate a power reference calculation flow for LSTM;

[0069] Figure 5 This is the overall flowchart of the present invention. Detailed Implementation

[0070] In order to achieve the purpose of this application and make the technical solution and advantages clearer, the following describes this application in further detail with reference to the accompanying drawings and embodiments.

[0071] This invention discloses an energy management method for electric vehicle hybrid energy storage systems incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm, comprising the following steps:

[0072] Step 1: Obtain electric vehicle driving cycle data, including vehicle speed V and load power P. l .

[0073] Multiple acceleration, constant speed, and deceleration intervals were set up, with acceleration intervals of 1–15s and 48–55s, constant speed intervals of 15–20s and 55–60s, and deceleration intervals of 20–30s and 60–65s, to simulate car driving and calculate the electrical load P required for driving at each time point. l The details are as follows:

[0074]

[0075] Among them, F m As the driving force, Wtg F is the tangent component of gravity (assuming a slope of 0). air For air resistance, F roll For rolling resistance, P m The mechanical power at the wheel is ρ, where ρ is the air density and C is the density of air. d Let A be the drag coefficient. f Where η is the vehicle's frontal area, η is the electromechanical conversion power, V is the driving speed, and C is the vehicle's frontal area. rr This is the tire deformation dissipation coefficient.

[0076] Step 2: Construct a multi-layer LSTM neural network controller, and process the historical state data (velocity V, load power P) l Supercapacitor voltage U c DC bus voltage error As input, the time-dependent features of driving conditions are learned, and the multi-layer LSTM neural network controller learns the mapping from the input state to power allocation, outputting the battery power reference component. and supercapacitor power reference component

[0077] The multilayer LSTM neural network controller adopts a multilayer structure, including an input layer V and a P layer. l U c and 50-cell LSTM layer, 40-cell fully connected layer and output layer, outputting battery power reference component. and supercapacitor power reference component

[0078] The multi-layer LSTM neural network controller learns the time-dependent features of driving conditions, as detailed below:

[0079] LSTM solves the dependency problem through a gating mechanism. The core formula is as follows:

[0080] Input gate: i t =σ(W i [h t-1 ,x t ]+b i )

[0081] Forgotten Gate: f t =σ(W f [h t-1 ,x t ]+b f )

[0082] Candidate memories:

[0083] Memory update:

[0084] Output gate: o t =σ(W o [h t-1 ,x t ]+b o )

[0085] Hidden layer output: h t =o t tanh(C t )

[0086] h is the current input. t-1 The hidden state is the state at the previous time step, ⊙ represents element-wise multiplication, σ and tanh are the sigmoid activation functions, and W is the hidden state at the previous time step. i W f W c W o These are the weights for the input gate, forget gate, candidate memory, and output gate, respectively, b. i ,b f ,b c ,b o These are the input gate, forget gate, candidate memory, and output gate biases, respectively, n. h ,n x To hide the number of units and the dimension of the input vector, a state persistence mechanism is used to retain historical state information to solve the temporal dependency problem. The state update equation is as follows:

[0087] ctht=ft ct-1+it ct~=ot tanh(ct) (8)

[0088] Where: ft: forget gate output, it: input gate output, ct~: candidate cell state, trained through LSTM layers and fully connected layers, finally outputting the battery power reference component. and supercapacitor power reference component

[0089] Step 3: Construct a fractional-order integral derivative FOID controller. The FOID controller uses the DC bus reference voltage. With the measured voltage U DC error As input, the output power correction component P FOID , will P FOID Decomposed into battery power reference components and supercapacitor power reference component Used to correct the power distribution results of the multilayer LSTM neural network controller, controlling the bus voltage fluctuation within ±3% of the reference value.

[0090] The FOID controller is as follows:

[0091] Fractional operator s of FOID controller λ Discretization was achieved using a 5th-order Oustaloup filter, and integral operators H were constructed respectively. i (s), corresponding to s -λ Differential operator H d (s), corresponding to s 1-λ :

[0092]

[0093] Among them, T z = 1ms is the sampling period, ω k ,ω' k For the recursive calculation of poles / zeros, K is the amplitude compensation gain, optimized through zero-pole matching, ω l ω is the low-frequency cutoff frequency. h This is the high-frequency cutoff frequency;

[0094] Input DC bus voltage With the measured voltage U DC error

[0095] Output power correction component:

[0096]

[0097] P FOID Decomposed into battery power reference component and supercapacitor power reference component:

[0098] Among them, K B and K C Used to adjust the dominance of the PID controller in power distribution between the battery and the supercapacitor, E vdc (s) is the Laplace transform of the voltage error, K i K is the integral gain, used for the integral term of the error, and is used to eliminate steady-state error. d The differential gain, used for the differential term of the error, is used for constant instantaneous fluctuations, L -1 λ is the inverse Laplace transform operator, and λ is the fractional calculus intermediate order, with values ​​ranging from [0.1, 0.9].

[0099] Step 4: Based on the optimization algorithm, a multi-objective joint optimization framework is adopted. The primary objective is to minimize the battery root mean square current, and the weights and biases of the multilayer LSTM neural network controller in Step 2 are optimized and trained. The secondary objective is to minimize the voltage error, and the FOID controller parameters (K) are optimized accordingly. i ,K d Optimization of λ); the objective function is:

[0100] Fitness = α·Ibatt,RMS +β·|E vdc | (1)

[0101] in, I bat (t) represents the battery current time series, T is the simulation time, α and β are the weighting coefficients of the primary and secondary objectives, α + β = 1, |E vdc | represents the Laplace transform of the voltage error, λ∈[0.1,0.9].

[0102] The improved fusion particle swarm optimization algorithm is used as follows:

[0103] Step 4.1: Initialize PSO algorithm parameters, including the number of particles / individuals n, the maximum number of iterations iter, the inertia weight u, the initial inertia u0, and the final inertia value u. f Set the catastrophic perturbation probability and the proportion of elite particles retained in the HO algorithm, with the upper and lower bounds of the particle position in each dimension being x. max x min ;

[0104] Step 4.2: The position of each particle / individual represents a set of weights and biases of the multilayer LSTM neural network controller with respect to FOID K. i ,K d ,λ; Hippo particle swarm fusion formula initialization of particle swarm:

[0105] x t =x min +r[(rand*x max -0.5)-x min (3)

[0106] Where r is a random number between 0 and 1, and rand is a random number;

[0107] The nonlinear decaying inertia weight is:

[0108] u = u0 - (u0-u f )*iter / n (4)

[0109] The formula for updating the position of each particle / individual is:

[0110] x iter = u*x0 + c1*r1*(x b -x0) + c2*r2*( x g - x0 ) (5)

[0111] Where, x b For the optimal value of the particle, x gThe global optimal value is represented by c1 and c2, which are acceleration factors, and r1 and r2 are social cognition coefficients.

[0112] Step 4.3: Compare the fitness obtained from the electrical simulation with the fitness corresponding to the best position in the particle's history, and set the calculation benchmark. If the particle gets stuck, use the HO algorithm to perform a local search and perturb the non-elite particles. If the particle gets stuck multiple times, keep the elite particles and reset the rest.

[0113] In the HO algorithm, elite particles represent male hippos, and the growth mechanism for non-elite particles is improved by introducing a new convergence factor T.

[0114]

[0115] Where, x t+1 This indicates the position after the HO algorithm performs a local search, t is the current iteration number, iter is the maximum iteration number, A is a random number between 0 and 1, and B is a random number between 1 and 2. When T > 0.9, the naive hippo tends to perform a fine search (local development) near the optimal solution, thus improving the algorithm's accuracy.

[0116] Step 4.4: For each particle / individual position, decode into a set of weights and biases of a multilayer LSTM neural network controller, and the parameters K of FOID. i ,K d ,λ;

[0117] Step 4.5: Perform electrical simulation using a set of weights and biases of a multilayer LSTM neural network controller under acceleration / deceleration driving cycles or target driving cycles, and use the decoded FOID parameter K. i ,K d ,λ is used to obtain the Laplace transform of the voltage error in the electrical simulation;

[0118] Step 4.6: Calculate the objective function value as the fitness;

[0119] Step 4.7: Update the optimal position and global optimal position of the particle / individual.

[0120] Step 4.8: Repeat steps 4.2-4.7 until the maximum number of iterations is reached or convergence is achieved. Output a set of weights and biases of the multilayer LSTM neural network controller corresponding to the global optimal position, as well as the parameter K of FOID. i ,K d ,λ.

[0121] Step 5: Convert the power reference component obtained in Step 2 With the power reference component from step 3 Dynamic weighted summation yields the final battery power reference value. and supercapacitor power reference value And input to the battery and supercapacitor DC / DC converter.

[0122] The output components of the FOID controller ( and ) respectively through the weight constant (K B ,K C After scaling, it participates in the dynamic weighted summation:

[0123]

[0124] Among them, dynamic weights Sensitivity coefficient k = 5.0, when |e vdc When |>2V, η(t)>0.88, FOID dominates voltage regulation.

[0125] The above method and system will be experimentally verified below:

[0126] Figure 1 To simulate the two acceleration and deceleration events of an electric vehicle, multiple acceleration and constant speed deceleration intervals were set: acceleration intervals of 1–15s and 48–55s, constant speed intervals of 15–20s and 55–60s, and deceleration intervals of 20–30s and 60–65s, mimicking the operation of an electric vehicle. The required electrical load P was calculated using the gradient integration method in MATLAB. load .

[0127] The required electrical load P load As a result, the Hippo Particle Swarm Optimization method uses the battery root mean square current as the primary objective and the voltage error as the secondary objective, as the input layer of the multi-layer LSTM for training. It optimizes the LSTM weights, bias, and FIOD controller parameters, and then uses the LSTM controller and FIOD controller weighted sum to generate the final power reference for the battery and supercapacitor.

[0128] Depend on Figure 3 It is evident that during rapid acceleration and deceleration, the FOID-dominated mode is activated through a dynamic weighting mechanism, enabling the supercapacitor to instantaneously bear 82% of the peak power while limiting the battery power to 9.5kW, effectively reducing the battery peak current to 105A. Combined with the timing prediction capability of LSTM, voltage fluctuations are stabilized within ±2.3%, and power switching overshoot is reduced to 4.3%, verifying the significant advantages of the fusion algorithm in the synergistic optimization of battery stress minimization and voltage stability.

[0129] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An energy management method for electric vehicle hybrid energy storage systems incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm, characterized in that, Includes the following steps: Step 1: Obtain electric vehicle driving cycle data, including vehicle speed V and load power P. l ; Step 2: Construct a multi-layer LSTM neural network controller, and process the historical state data (velocity V, load power P) l Supercapacitor voltage U c DC bus voltage error As input, the time-dependent features of driving conditions are learned, and the multi-layer LSTM neural network controller learns the mapping from the input state to power allocation, outputting the battery power reference component. and supercapacitor power reference component Step 3: Construct a fractional-order integral derivative FOID controller. The FOID controller uses the DC bus reference voltage. With the measured voltage U DC error As input, the output power correction component P FOID , will P FOID Decomposed into battery power reference components and supercapacitor power reference component Used to correct the power distribution results of the multilayer LSTM neural network controller, and control the bus voltage fluctuation within ±3% of the reference value; Step 4: Based on the optimization algorithm, a multi-objective joint optimization framework is adopted. The primary objective is to minimize the battery root mean square current, and the weights and biases of the multilayer LSTM neural network controller in Step 2 are optimized and trained. The secondary objective is to minimize the voltage error, and the FOID controller parameters (K) are optimized accordingly. i ,K d Optimization of λ); the multi-objective optimization function is: Fitness=α·I batt,RMS +β·|E vdc |(1) in, I bat (t) represents the battery current time series, T is the simulation time, α and β are the weighting coefficients of the primary and secondary objectives, α + β = 1, |E vdc | represents the Laplace transform of the voltage error, λ∈[0.1,0.9]; Step 5: Convert the power reference component obtained in Step 2 With the power reference component from step 3 Dynamic weighted summation yields the final battery power reference value. and supercapacitor power reference value And input to the battery and supercapacitor DC / DC converter.

2. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 1, characterized in that, In step 1, the load power P l The calculation method is as follows: Multiple acceleration, constant speed, and deceleration intervals were set up, with acceleration intervals of 1–15s and 48–55s, constant speed intervals of 15–20s and 55–60s, and deceleration intervals of 20–30s and 60–65s, to simulate car driving and calculate the electrical load P required for driving at each time point. l The details are as follows: Among them, F m As the driving force, W tg F is the tangent component of gravity (assuming a slope of 0). air For air resistance, F roll For rolling resistance, P m The mechanical power at the wheel is ρ, where ρ is the air density and C is the density of air. d Let A be the drag coefficient. f Where η is the vehicle's frontal area, η is the electromechanical conversion power, V is the driving speed, and C is the vehicle's frontal area. rr This is the tire deformation dissipation coefficient.

3. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 1, characterized in that, Step 4 utilizes an improved fusion particle swarm optimization algorithm, as detailed below: Step 4.1: Initialize PSO algorithm parameters, including the number of particles / individuals n, the maximum number of iterations iter, the inertia weight u, the initial inertia u0, and the final inertia value u. f Set the catastrophic perturbation probability and the proportion of elite particles retained in the HO algorithm, with the upper and lower bounds of the particle position in each dimension being x. max x min ; Step 4.2: The position of each particle / individual represents a set of weights and biases of the multilayer LSTM neural network controller with respect to FOID K. i ,K d ,λ; Hippo particle swarm fusion formula initialization of particle swarm: x t =x min +r[(rand*x max -0.5)-x min ](3) Where r is a random number between 0 and 1, and rand is a random number; The nonlinear decaying inertia weight is: u=u0-(u0-u f )*iter / n(4) The formula for updating the position of each particle / individual is: x iter =u*x0+c1*r1*(x b -x0)+c2*r2*(x g -x0)(5) Where, x b For the optimal value of the particle, x g The global optimal value is represented by c1 and c2, which are acceleration factors, and r1 and r2 are social cognition coefficients. Step 4.3: Compare the fitness obtained from the electrical simulation with the fitness corresponding to the best position in the particle's history, and set the calculation benchmark. If the particle gets stuck, use the HO algorithm to perform a local search and perturb the non-elite particles. If the particle gets stuck multiple times, keep the elite particles and reset the rest. In the HO algorithm, elite particles represent male hippos, and the growth mechanism for non-elite particles is improved by introducing a new convergence factor T. Where, x t+1 This indicates the position after the HO algorithm performs a local search, t is the current iteration number, iter is the maximum iteration number, A is a random number between 0 and 1, and B is a random number between 1 and 2. When T > 0.9, the naive hippo tends to perform a fine search (local development) near the optimal solution, thus improving the algorithm's accuracy. Step 4.4: For each particle / individual position, decode into a set of weights and biases of a multilayer LSTM neural network controller, and the parameters K of FOID. i ,K d ,λ; Step 4.5: Perform electrical simulation using a set of weights and biases of a multilayer LSTM neural network controller under acceleration / deceleration driving cycles or target driving cycles, and use the decoded FOID parameter K. i ,K d ,λ is used to obtain the Laplace transform of the voltage error in the electrical simulation; Step 4.6: Calculate the objective function value as the fitness; Step 4.7: Update the optimal position and global optimal position of the particle / individual. Step 4.8: Repeat steps 4.2-4.7 until the maximum number of iterations is reached or convergence is achieved. Output a set of weights and biases of the multilayer LSTM neural network controller corresponding to the global optimal position, as well as the parameter K of FOID. i ,K d ,λ.

4. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 1, characterized in that, The multi-layer LSTM neural network controller adopts a multi-layer structure, including input layers V and P. l U c and 50-cell LSTM layer, 40-cell fully connected layer and output layer, outputting battery power reference component. and supercapacitor power reference component 5. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 4, characterized in that, The multi-layer LSTM neural network controller learns the time-dependent features of driving conditions, as detailed below: LSTM solves the dependency problem through a gating mechanism. The core formula is as follows: Input gate: i t =σ(W i [h t-1 ,x t ]+b i ) Forgotten Gate: f t =σ(W f [h t-1 ,x t ]+b f ) Candidate memories: Memory update: Output gate: o t =σ(W o [h t-1 ,x t ]+b o ) Hidden layer output: h t =o t tanh(C t ) h is the current input. t-1 The hidden state is the state at the previous time step, ⊙ represents element-wise multiplication, σ and tanh are the sigmoid activation functions, and W is the hidden state at the previous time step. i W f W c W o These are the weights for the input gate, forget gate, candidate memory, and output gate, respectively, b. i ,b f ,b c ,b o These are the input gate, forget gate, candidate memory, and output gate biases, respectively, n. h ,n x To hide the number of units and the dimension of the input vector, a state persistence mechanism is used to retain historical state information to solve the temporal dependency problem. The state update equation is as follows: ctht=ft ct-1+it ct~=ot tanh(ct)(8) Where: ft: forget gate output, it: input gate output, ct~: candidate cell state, trained through LSTM layers and fully connected layers, finally outputting the battery power reference component. and supercapacitor power reference component 6. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 1, characterized in that, The FOID controller in step 3 is specifically as follows: Fractional operator s of FOID controller λ Discretization was achieved using a 5th-order Oustaloup filter, and integral operators H were constructed respectively. i (s), corresponding to s -λ Differential operator H d (s), corresponding to s 1-λ : H i (z)≈s -λ ,H d (z)≈s 1-λ Among them, T z = 1ms is the sampling period, ω k ,ω ' k For the recursive calculation of poles / zeros, K is the amplitude compensation gain, optimized through zero-pole matching, ω l ω is the low-frequency cutoff frequency. h This is the high-frequency cutoff frequency; Input DC bus voltage With the measured voltage U DC error Output power correction component: P FOID Decomposed into battery power reference component and supercapacitor power reference component: Among them, K B and K C Used to adjust the dominance of the PID controller in power distribution between the battery and the supercapacitor, E vdc (s) is the Laplace transform of the voltage error, K i K is the integral gain, used for the integral term of the error, and is used to eliminate steady-state error. d The differential gain, used for the differential term of the error, is used for constant instantaneous fluctuations, L -1 λ is the inverse Laplace transform operator, and λ is the fractional calculus intermediate order, with values ​​ranging from [0.1, 0.9].

7. The energy management method for a hybrid energy storage system for electric vehicles incorporating the Hippodrome Particle Swarm Optimization (LSTM) algorithm as described in claim 1, characterized in that, The output component of the FOID controller ( and ) respectively through the weight constant (K B ,K C After scaling, it participates in the dynamic weighted summation: Among them, dynamic weights The sensitivity coefficient k = 5.0, when |E vdc When |>2V, η(t)>0.88, FOID dominates voltage regulation.

Images