Electric vehicle energy management method and system based on lithium-sodium hybrid battery system
By optimizing the energy management of lithium-sodium hybrid batteries using deep Q-networks and lightweight time-series models, coordination between power and thermal management is achieved, thereby improving the driving range and energy utilization efficiency of electric vehicles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV OF TECH
- Filing Date
- 2026-03-26
- Publication Date
- 2026-05-12
AI Technical Summary
Existing lithium-sodium hybrid battery systems cannot effectively coordinate power distribution and thermal management, resulting in low energy utilization efficiency and failure to maximize driving range.
The initial optimal control strategy is generated by training a deep Q-network, and then optimized by combining a lightweight time series model and an augmented Lagrangian function to achieve power balance and thermal safety coordination of the lithium-sodium hybrid battery.
It significantly improves energy utilization efficiency, maximizes driving range, and meets real-time energy management requirements, solving the problem that traditional strategies cannot adaptively adjust power.
Smart Images
Figure CN122008950A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of energy storage system technology, and in particular to an energy management method and system for electric vehicles based on a lithium-sodium hybrid battery system. Background Technology
[0002] With increasing global emphasis on environmental protection, electric vehicles (EVs) are gaining more attention due to their zero emissions and low-noise operation. However, in low-temperature environments, the driving range of EVs decreases significantly, mainly due to degraded battery performance, increased thermal management energy consumption, and reduced system efficiency. Lithium-ion batteries experience increased internal resistance and a substantial decrease in usable capacity at low temperatures, while sodium-ion batteries exhibit excellent low-temperature performance but have lower energy density. Lithium-sodium hybrid batteries, as a novel energy storage technology combining the advantages of lithium-ion and sodium-ion batteries, have made significant progress in recent years in terms of technological synergy, application expansion, and industrialization. However, existing energy management strategies often fail to effectively coordinate power distribution and thermal management in lithium-sodium hybrid batteries, resulting in low energy utilization efficiency and an inability to maximize driving range. Summary of the Invention
[0003] In view of this, the present invention proposes an energy management method and system for electric vehicles based on a lithium-sodium hybrid battery system.
[0004] The technical solution of the present invention is implemented as follows: The first aspect of the present invention provides an energy management method for electric vehicles based on a lithium-sodium hybrid battery system, comprising: State transition data of lithium-sodium hybrid battery samples are acquired, and a preset deep Q-network is trained using the state transition data to generate an initial optimal control strategy. The state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition. The initial optimal control strategy includes the initial output power and the initial load power. The lightweight time series model is processed by knowledge distillation, structural pruning and linear quantization to obtain the target TTM prediction model. The power feature sequence of the lithium-sodium hybrid battery sample is input into the target TTM prediction model to obtain future power prediction information. Using minimizing the initial output power and the initial load power as the objective function, and embedding power balance constraints into the objective function, an augmented Lagrangian function is constructed. The future power prediction information is input into the augmented Lagrangian function, and the initial output power and the initial load power in the initial optimal control strategy are alternately optimized to obtain the target optimal control strategy.
[0005] Based on the above technical solutions, preferably, the preset depth Q-network includes a target network and a prediction network; the step of training the preset depth Q-network using the state transition data to generate an initial optimal control policy includes: The state transition data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value; The mean squared error loss function is calculated based on the target Q value and the current Q value, and the weights of the target network are iteratively updated by backpropagation using the mean squared error loss function until the mean squared error loss function converges, thus obtaining the target depth Q network. The target depth Q-network is used to traverse the state transition data, filter out the target execution action corresponding to the maximum target Q value, and determine the initial output power and initial load power corresponding to the target execution action as the initial optimal control strategy.
[0006] Based on the above technical solution, preferably, before inputting the state transition data into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value, the method further includes: The sampling priority of each group of state transition data is determined based on the timing difference error of each group of state transition data; the larger the timing difference error of the state transition data, the higher the sampling priority. By sampling from the state transition data according to the sampling priority, training sample data is obtained, and the training sample data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value.
[0007] Based on the above technical solutions, preferably, the step of performing lightweight processing on the lightweight time series model through knowledge distillation, structural pruning, and linear quantization to obtain the target TTM prediction model includes: The lightweight time series model is trained using a large model as a teacher model, and the output difference is minimized to obtain the first TTM prediction model. Based on the scaling factor of the BN layer, structural pruning is performed on the channels in the first TTM prediction model whose contribution is lower than a preset threshold to obtain the second TTM prediction model. Linear quantization is used to linearly map the weights of the second TTM prediction model to an 8-bit integer range to obtain the target TTM prediction model.
[0008] Based on the above technical solutions, preferably, the step of constructing an augmented Lagrangian function with the objective function of minimizing the initial output power and the initial load power, and embedding power balance constraints into the objective function, includes: A dual-package cooperative optimization problem is constructed in the rolling time domain, with the objective function being to minimize the initial output power and the initial load power. Power balance constraints are embedded in the objective function to construct an augmented Lagrangian function. The power balance constraints include power balance, power constraints, and SOC constraints for the initial output power and the initial load power.
[0009] Based on the above technical solutions, preferably, the step of inputting the future power prediction information into the augmented Lagrange function to alternately optimize the initial output power and the initial load power in the initial optimal control strategy to obtain the target optimal control strategy includes: While keeping the Lagrange multipliers constant, optimize the initial load power by fixing the initial output power, or optimize the initial output power by fixing the initial load power. The optimal control strategy for the target is obtained by iteratively updating the coordinates in reverse cyclic descent.
[0010] Based on the above technical solutions, preferably, the step of iteratively updating the target optimal control strategy using the reverse cyclic coordinate descent method further includes: After the power changes of the initial output power and the initial load power are less than a preset power threshold, the Lagrange multipliers and penalty factors in the augmented Lagrange function are updated to balance the constraint error and the convergence speed.
[0011] More preferably, a second aspect of the present invention provides an electric vehicle energy management system based on a lithium-sodium hybrid battery system, comprising: an initial strategy generation module, a future power prediction module, and a target strategy generation module; wherein, The initial strategy generation module is configured to acquire state transition data of lithium-sodium hybrid battery samples and use the state transition data to train a preset deep Q-network to generate an initial optimal control strategy; the state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition; the initial optimal control strategy includes the initial output power and the initial load power; The future power prediction module is configured to perform lightweight processing on the lightweight time series model through knowledge distillation, structural pruning and linear quantization to obtain the target TTM prediction model, and input the power feature sequence of the lithium-sodium hybrid battery sample into the target TTM prediction model to obtain future power prediction information. The target strategy generation module is configured to use minimizing the initial output power and the initial load power as the objective function, embed power balance constraints into the objective function, construct an augmented Lagrangian function, input the future power prediction information into the augmented Lagrangian function, and alternately optimize the initial output power and the initial load power in the initial optimal control strategy to obtain the target optimal control strategy.
[0012] More preferably, a third aspect of the present invention provides an electronic device, including a processor and a memory; the memory has a computer program stored thereon, wherein the computer program, when executed by the processor, implements the electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in the first aspect.
[0013] More preferably, a fourth aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in the first aspect.
[0014] The electric vehicle energy management method and system based on a lithium-sodium hybrid battery system of the present invention have the following advantages over the prior art: 1. By collecting the current state, actions, rewards, and new states of the lithium-sodium hybrid battery, a dynamic environment model is constructed. An initial optimal power strategy is generated using reinforcement learning, and iterative optimization is performed in a simulation environment to ensure the robustness of the strategy. Based on this, a lightweight time-series model is used to accurately predict future power, and augmented Lagrange alternation optimization is combined to achieve global optimum under constraints. This enables coordinated control of the lithium-sodium hybrid battery's output power and load power, effectively balancing battery power distribution and thermal safety boundaries, significantly improving energy utilization efficiency, maximizing driving range, and achieving low-computing-power, low-latency deployment through model lightweighting, meeting real-time energy management requirements.
[0015] 2. By automatically learning the nonlinear mapping relationship between the state, action, and reward of lithium-sodium hybrid batteries, the system accurately captures the differences in dynamic characteristics between lithium and sodium batteries, solving the problem that traditional strategies cannot adaptively adjust power. A lightweight, high-precision time-series prediction model is constructed to achieve rapid and accurate prediction of future power, proactively sensing future power fluctuations and load changes, transforming energy management from passive response to proactive control, and avoiding thermal runaway caused by sudden power changes.
[0016] 3. By embedding power balance constraints, battery safety constraints, and thermal management constraints into the optimization objectives, the power coordination and matching between lithium-ion batteries and sodium-ion batteries are strictly guaranteed. Alternating iterative optimization of output power and load power achieves globally optimal synergy in the lithium-sodium hybrid battery, resolving the pain point of ineffective power allocation coordination. Based on this, and combined with future power prediction information, a globally optimal, constraint-feasible, and thermally safe target control strategy is achieved. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating an electric vehicle energy management method based on a lithium-sodium hybrid battery system, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an electric vehicle energy management system based on a lithium-sodium hybrid battery system, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0020] In some embodiments, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating an electric vehicle energy management method based on a lithium-sodium hybrid battery system, provided by an embodiment of the present invention. The electric vehicle energy management method based on a lithium-sodium hybrid battery system provided by the present invention includes: S110: Acquire state transition data of lithium-sodium hybrid battery samples, and use the state transition data to train a preset deep Q-network to generate an initial optimal control strategy; the state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition; the initial optimal control strategy includes the initial output power and the initial load power.
[0021] In this embodiment, the lithium-sodium hybrid battery sample combines a lithium battery and a sodium battery as power sources. The sodium battery does not require heating, while the lithium battery does require heating to maintain output. The lithium-sodium hybrid battery sample is connected to the load via a DC-DC converter and manages the battery output according to power demand and SOC.
[0022] In some embodiments, the preset deep Q-network includes a target network and a prediction network; training the preset deep Q-network using state transition data to generate an initial optimal control policy includes: The state transition data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and the current Q value; The mean squared error loss function is calculated based on the target Q value and the current Q value. The weights of the target network are then iteratively updated using backpropagation based on the mean squared error loss function until the mean squared error loss function converges, thus obtaining the target depth Q network. By using a target depth Q-network to traverse the state transition data, the target execution action corresponding to the maximum target Q value is selected, and the initial output power and initial load power corresponding to the target execution action are determined as the initial optimal control strategy.
[0023] In this embodiment, an experience replay mechanism is employed to store state transition data generated during the experiment, including the current state, the action performed, the reward obtained, and the new state after the transition. A target network structure is introduced, and the target Q-value is calculated by maintaining a target network independent of the prediction network. By combining double Q-learning, action selection and Q-value evaluation are decoupled, suppressing the possibility of overestimating the Q-value in Q-learning.
[0024] In some embodiments, before inputting the state transition data into the target network and the prediction network respectively to obtain the corresponding target Q-value and current Q-value, the method further includes: The sampling priority of each set of state transition data is determined based on the timing difference error of each set of state transition data; the larger the timing difference error of the state transition data, the higher the sampling priority. By combining sampling priorities, sample data is obtained from the state transition data. The training sample data is then input into the target network and the prediction network respectively to obtain the corresponding target Q value and the current Q value.
[0025] In this embodiment, different sampling priorities are assigned to each set of state transition data based on its temporal differential error (TD-error). State transition data with larger errors are more likely to be selected for training, thereby accelerating the learning of key experiences and improving the overall learning efficiency and convergence speed of the algorithm. After completing the above training process, the system finally generates an optimal control lookup table covering all possible operating conditions. This lookup table encapsulates the initial optimal control strategy obtained through training.
[0026] S120 performs lightweight processing on the lightweight time series model through knowledge distillation, structural pruning, and linear quantization to obtain the target TTM prediction model. The power feature sequence of the lithium-sodium hybrid battery sample is input into the target TTM prediction model to obtain future power prediction information.
[0027] In some embodiments, a lightweight time series model is lightly processed through knowledge distillation, structural pruning, and linear quantization to obtain a target TTM prediction model, including: The first TTM prediction model is obtained by using a large model as a teacher model to train a lightweight time series model and minimizing the output difference. Based on the scaling factor of the BN layer, structural pruning is performed on channels whose contributions in the first TTM prediction model are lower than a preset threshold to obtain the second TTM prediction model. Linear quantization is used to linearly map the weights of the second TTM prediction model to an 8-bit integer range to obtain the target TTM prediction model.
[0028] In this embodiment, TTM stands for Tiny Time Mixers, and BN stands for Batch Normalization. The sub-model structure of the lightweight time series model includes a multi-scale time series mixing layer and a readout layer. The multi-scale time series mixing layer can be represented as:
[0029] ; in, Characterize the input feature sequence, such as a time series or time-series feature map. Depth-separable convolutions can be represented. Different kernel sizes / stride sizes are used to capture local features at different time scales. Characterize the pooling layer, This is the pooling window size, used for downsampling and extracting global temporal features. Feature concatenation involves concatenating outputs from different scales along the channel dimension to obtain multi-scale fused features. , Characterizing global average pooling, for Perform global pooling to compress the spatial / temporal dimensions and obtain a channel-level global description. The Sigmoid activation function maps the output to... The interval, as the gate weight, Here are the weight matrix and biases for the gated fully connected layer, used to learn adaptive gated weights. This is the gating weight vector, which controls the feature fusion ratio of the subsequent two branches. Characterized by element-wise multiplication Characterizing pointwise convolution, Characterizing feedforward neural networks, Characterizes the output features after mixing.
[0030] The readout layer can be represented as: ; in, Layer normalization, for Normalization helps stabilize training and accelerate convergence. Characterize the prediction head, Characterization from time arrive The predicted sequence, It predicts the step size.
[0031] The loss function during training can be expressed as: ; in, To regress the loss, For time step index, the range of values is... ; Representation at time For the future Predicted value at time, Characterized in The actual observed value at that moment, For residuals; Let be the threshold hyperparameter for Huber loss. Here is the Huber loss function.
[0032] The physical penalty term can be represented as: ; The total loss is: ; in, This is a physical constraint penalty term. The weighting coefficient for the penalty term. For a moment Predicted power / physical quantity For a moment The maximum permissible physical power.
[0033] Knowledge distillation, which uses a large model as the teacher network, minimizes output discrepancies: ; in, For knowledge distillation loss, This is the distillation loss weighting coefficient. Representation of student models in Always Predicted output at time step Representation of teacher models in Always Predicted output at time step.
[0034] Structural pruning is based on channel importance, such as using the scaling factor in BN layers to prune low-contribution channels to meet computational budget requirements. ; in, Representing floating-point operands, This is the pruned model function. These are the parameters of the pruned model. The FLOPs Budget is the maximum amount of computation allowed by the onboard equipment.
[0035] Linear quantization can be expressed as: ; in, The weights are quantized 8-bit integers. For the original floating-point weights, As a proportion, It is zero point.
[0036] Using QAT can further reduce accuracy loss, thereby unifying the objectives, i.e., the model for each working condition: ; in, Characterizing the model parameters after pruning Perform minimization optimization. This is an abbreviation for constraint conditions. Characterizes the magnitude of the parameters of the model after pruning. Characterize memory budget, Representation model inference delay Represents real-time budgeting.
[0037] S130 uses minimizing the initial output power and initial load power as the objective function, embeds the power balance constraint into the objective function, constructs an augmented Lagrangian function, inputs future power prediction information into the augmented Lagrangian function, and alternately optimizes the initial output power and initial load power in the initial optimal control strategy to obtain the target optimal control strategy.
[0038] In some embodiments, an augmented Lagrangian function is constructed by minimizing the initial output power and the initial load power as the objective function, and embedding power balance constraints into the objective function, including: A dual-package cooperative optimization problem is constructed in the rolling time domain, with the objective function being to minimize the initial output power and the initial load power. Power balance constraints are embedded in the objective function to construct an augmented Lagrangian function. The power balance constraints include power balance, power constraints, and SOC constraints for the initial output power and the initial load power.
[0039] In this embodiment, in the rolling time domain The above constructs a dual-package collaborative optimization problem, and the objective function can be expressed as: ; The power balance is: ; The power constraint is: ; The SOC constraint is: ; in, Characterization with and To optimize variables and minimize the objective function , For time step index, These are the weighting coefficients. Characterizing the initial output power, Characterizing the initial load power, Characterizes the predicted total power demand.
[0040] The augmented Lagrangian function can be expressed as: ; in, These are Lagrange multipliers, i.e., dual variables. , which is a penalty factor used to adjust the convergence rate.
[0041] By using the augmented Lagrangian function, the original constrained optimization is transformed into a form of "unconstrained + augmented penalty term", creating conditions for subsequent alternating solutions.
[0042] In some embodiments, future power prediction information is input into the augmented Lagrangian function to alternately optimize the initial output power and initial load power in the initial optimal control strategy, thereby obtaining the target optimal control strategy, including: While keeping the Lagrange multipliers constant, optimize the initial load power while keeping the initial output power constant, or optimize the initial output power while keeping the initial load power constant. The optimal control strategy for the target is obtained by iteratively updating the coordinates in reverse cyclic descent.
[0043] In this embodiment, fixed Then, the initial optimal control strategy is... Alternating optimization is performed. Since each subproblem is a one-dimensional convex quadratic problem, it can be solved quickly or iterated efficiently.
[0044] fixed ,renew : ; fixed ,renew : ; The "Reverse Circular Coordinate Descent (RCCD)" method is used for iterative updates: the variables that contribute the most to the objective function are updated first, and the variables are traversed in reverse loop to avoid local oscillations and repeated convergence paths. Each update step can be completed in a closed-form solution or iteratively with very low complexity.
[0045] In some embodiments, the optimal control strategy is obtained by iteratively updating the target using a reverse cyclic coordinate descent method, and further includes: Once the power changes in the initial output power and initial load power are less than the preset power threshold, the Lagrange multipliers and penalty factors in the augmented Lagrange function are updated to balance the constraint error and convergence speed.
[0046] After the initial optimal control policy converges, the Lagrange multipliers and penalty factors are updated to strengthen the constraint convergence: ; If the constraint error is large, the penalty factor can be increased adaptively. To accelerate convergence If the error converges too quickly, the penalty factor can be reduced. Prevent shocks.
[0047] After completing one round of solving, only the optimal solution at the current time step is applied: ; Then the state is moved forward, the window is updated, and we proceed to the next step of prediction and optimization: .
[0048] In some embodiments, please refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of an electric vehicle energy management system based on a lithium-sodium hybrid battery system, provided as an embodiment of the present invention. The present invention provides an electric vehicle energy management system 200 based on a lithium-sodium hybrid battery system, comprising: an initial strategy generation module 210, a future power prediction module 220, and a target strategy generation module 230; wherein,
[0049] The initial strategy generation module 210 is configured to acquire state transition data of lithium-sodium hybrid battery samples and use the state transition data to train a preset deep Q-network to generate an initial optimal control strategy; the state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition; the initial optimal control strategy includes the initial output power and the initial load power; The future power prediction module 220 is configured to perform lightweight processing on the lightweight time series model through knowledge distillation, structural pruning and linear quantization to obtain the target TTM prediction model. The power feature sequence of the lithium-sodium hybrid battery sample is input into the target TTM prediction model to obtain future power prediction information. The target strategy generation module 230 is configured to minimize the initial output power and the initial load power as the objective function, embed the power balance constraint into the objective function, construct an augmented Lagrangian function, input future power prediction information into the augmented Lagrangian function, and alternately optimize the initial output power and the initial load power in the initial optimal control strategy to obtain the target optimal control strategy.
[0050] In some embodiments, the preset deep Q-network includes a target network and a prediction network; the initial policy generation module 210 is specifically configured as follows: The state transition data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and the current Q value; The mean squared error loss function is calculated based on the target Q value and the current Q value. The weights of the target network are then iteratively updated using backpropagation based on the mean squared error loss function until the mean squared error loss function converges, thus obtaining the target depth Q network. By using a target depth Q-network to traverse the state transition data, the target execution action corresponding to the maximum target Q value is selected, and the initial output power and initial load power corresponding to the target execution action are determined as the initial optimal control strategy.
[0051] In some embodiments, the electric vehicle energy management system 200 based on a lithium-sodium hybrid battery system further includes a data sampling module; the data sampling module is specifically configured as follows: The sampling priority of each set of state transition data is determined based on the timing difference error of each set of state transition data; the larger the timing difference error of the state transition data, the higher the sampling priority. By combining sampling priorities, sample data is obtained from the state transition data. The training sample data is then input into the target network and the prediction network respectively to obtain the corresponding target Q value and the current Q value.
[0052] In some embodiments, the future power prediction module 220 is specifically configured as follows: The first TTM prediction model is obtained by using a large model as a teacher model to train a lightweight time series model and minimizing the output difference. Based on the scaling factor of the BN layer, structural pruning is performed on channels whose contributions in the first TTM prediction model are lower than a preset threshold to obtain the second TTM prediction model. Linear quantization is used to linearly map the weights of the second TTM prediction model to an 8-bit integer range to obtain the target TTM prediction model.
[0053] In some embodiments, the target policy generation module 230 is specifically configured as follows: A dual-package cooperative optimization problem is constructed in the rolling time domain, with the objective function being to minimize the initial output power and the initial load power. Power balance constraints are embedded in the objective function to construct an augmented Lagrangian function. The power balance constraints include power balance, power constraints, and SOC constraints for the initial output power and the initial load power.
[0054] In some embodiments, the target policy generation module 230 is specifically configured as follows: While keeping the Lagrange multipliers constant, optimize the initial load power while keeping the initial output power constant, or optimize the initial output power while keeping the initial load power constant. The optimal control strategy for the target is obtained by iteratively updating the coordinates in reverse cyclic descent.
[0055] In some embodiments, the target policy generation module 230 is specifically configured as follows: Once the power changes in the initial output power and initial load power are less than the preset power threshold, the Lagrange multipliers and penalty factors in the augmented Lagrange function are updated to balance the constraint error and convergence speed.
[0056] It should be noted that the electric vehicle energy management system based on the lithium-sodium hybrid battery system provided in this application embodiment and the electric vehicle energy management method based on the lithium-sodium hybrid battery system provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned electric vehicle energy management method based on the lithium-sodium hybrid battery system, and the repeated parts will not be described again.
[0057] In some embodiments, please refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 300 provided in this application includes a processor 310 and a memory 320; the memory 320 stores a computer program, wherein the computer program, when executed by the processor, implements the aforementioned electric vehicle energy management method based on a lithium-sodium hybrid battery system.
[0058] Specifically, processor 310 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. Processor 310 may also include onboard memory for caching purposes. Processor 310 may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.
[0059] The memory 320 may be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, the memory 320 may include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, apparatuses, or propagation media. Specific examples of the memory 320 include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and may also be random access memory (RAM) or flash memory; and / or wired / wireless communication links.
[0060] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described electric vehicle energy management method based on a lithium-sodium hybrid battery system. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.
[0061] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.
[0062] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.
Claims
1. An energy management method for electric vehicles based on a lithium-sodium hybrid battery system, characterized in that, include: State transition data of lithium-sodium hybrid battery samples are acquired, and a preset deep Q-network is trained using the state transition data to generate an initial optimal control strategy. The state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition. The initial optimal control strategy includes the initial output power and the initial load power. The lightweight time series model is processed by knowledge distillation, structural pruning and linear quantization to obtain the target TTM prediction model. The power feature sequence of the lithium-sodium hybrid battery sample is input into the target TTM prediction model to obtain future power prediction information. Using minimizing the initial output power and the initial load power as the objective function, and embedding power balance constraints into the objective function, an augmented Lagrangian function is constructed. The future power prediction information is input into the augmented Lagrangian function, and the initial output power and the initial load power in the initial optimal control strategy are alternately optimized to obtain the target optimal control strategy.
2. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 1, characterized in that, The preset depth Q-network includes a target network and a prediction network; The step of training a preset deep Q-network using the state transition data to generate an initial optimal control policy includes: The state transition data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value; The mean squared error loss function is calculated based on the target Q value and the current Q value, and the weights of the target network are iteratively updated by backpropagation using the mean squared error loss function until the mean squared error loss function converges, thus obtaining the target depth Q network. The target depth Q-network is used to traverse the state transition data, filter out the target execution action corresponding to the maximum target Q value, and determine the initial output power and initial load power corresponding to the target execution action as the initial optimal control strategy.
3. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 2, characterized in that, Before inputting the state transition data into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value, the method further includes: The sampling priority of each group of state transition data is determined based on the timing difference error of each group of state transition data; the larger the timing difference error of the state transition data, the higher the sampling priority. By sampling from the state transition data according to the sampling priority, training sample data is obtained, and the training sample data is input into the target network and the prediction network respectively to obtain the corresponding target Q value and current Q value.
4. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 1, characterized in that, The lightweight time series model is processed through knowledge distillation, structural pruning, and linear quantization to obtain the target TTM prediction model, including: The lightweight time series model is trained using a large model as a teacher model, and the output difference is minimized to obtain the first TTM prediction model. Based on the scaling factor of the BN layer, structural pruning is performed on the channels in the first TTM prediction model whose contribution is lower than a preset threshold to obtain the second TTM prediction model. Linear quantization is used to linearly map the weights of the second TTM prediction model to an 8-bit integer range to obtain the target TTM prediction model.
5. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 1, characterized in that, The process of constructing an augmented Lagrangian function with the objective function of minimizing the initial output power and the initial load power, and embedding power balance constraints into the objective function, includes: A dual-package cooperative optimization problem is constructed in the rolling time domain, with the objective function being to minimize the initial output power and the initial load power. Power balance constraints are embedded in the objective function to construct an augmented Lagrangian function. The power balance constraints include power balance, power constraints, and SOC constraints for the initial output power and the initial load power.
6. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 1, characterized in that, The step of inputting the future power prediction information into the augmented Lagrangian function to alternately optimize the initial output power and the initial load power in the initial optimal control strategy to obtain the target optimal control strategy includes: While keeping the Lagrange multipliers constant, optimize the initial load power by fixing the initial output power, or optimize the initial output power by fixing the initial load power. The optimal control strategy for the target is obtained by iteratively updating the coordinates in reverse cyclic descent.
7. The electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in claim 6, characterized in that, The method of iteratively updating the target optimal control strategy using the reverse cyclic coordinate descent method further includes: After the power changes of the initial output power and the initial load power are less than a preset power threshold, the Lagrange multipliers and penalty factors in the augmented Lagrange function are updated to balance the constraint error and the convergence speed.
8. An electric vehicle energy management system based on a lithium-sodium hybrid battery system, characterized in that, include: The system comprises an initial policy generation module, a future power prediction module, and a target policy generation module; among which, The initial strategy generation module is configured to acquire state transition data of lithium-sodium hybrid battery samples and use the state transition data to train a preset deep Q-network to generate an initial optimal control strategy; the state transition data includes the current state, the action performed, the reward obtained, and the new state after the transition; the initial optimal control strategy includes the initial output power and the initial load power; The future power prediction module is configured to perform lightweight processing on the lightweight time series model through knowledge distillation, structural pruning and linear quantization to obtain the target TTM prediction model, and input the power feature sequence of the lithium-sodium hybrid battery sample into the target TTM prediction model to obtain future power prediction information. The target strategy generation module is configured to use minimizing the initial output power and the initial load power as the objective function, embed power balance constraints into the objective function, construct an augmented Lagrangian function, input the future power prediction information into the augmented Lagrangian function, and alternately optimize the initial output power and the initial load power in the initial optimal control strategy to obtain the target optimal control strategy.
9. An electronic device comprising a processor and a memory; said memory storing a computer program, wherein, When executed by the processor, the computer program implements the electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the electric vehicle energy management method based on a lithium-sodium hybrid battery system as described in any one of claims 1 to 7.