Operation optimization method and system of new energy heavy truck battery power system

Abstract

The invention discloses an operation optimization method and system for a new energy heavy truck battery power system, and the method comprises the steps: inputting the current vehicle operation condition data of a new energy heavy truck into a pre-constructed decision network architecture, and outputting a state evaluation result and an action evaluation result; combining the state evaluation result with the action evaluation result to generate a plurality of candidate control strategies for operation optimization of the new energy heavy truck battery power system; the method comprises the following steps: performing feature extraction on multi-modal operation data from different sensors of a new energy heavy truck battery power system by adopting a self-supervised learning method to obtain unified feature representation; and performing matching evaluation on the unified feature representation and candidate control strategies, and selecting an optimal control strategy from a plurality of candidate control strategies to perform operation optimization of the new energy heavy truck battery power system. According to the invention, operation optimization can be accurately and efficiently carried out on the new energy heavy truck battery power system.

Description

Technical Field

[0001] This invention relates to the field of power battery technology, and in particular to an operation optimization method and system for a new energy heavy-duty truck battery power system. Background Technology

[0002] The core of optimizing the operation of the battery power system for new energy heavy-duty trucks lies in achieving efficient energy utilization and improved adaptability under multiple operating conditions through intelligent decision-making. Traditional approaches in existing technologies mainly rely on rule-based control (RBC) strategies or single-scale mechanistic models to drive control methods. For example, traditional energy management strategies use preset rules (such as fixed power allocation ratios or SOC thresholds) to statically control battery charging and discharging, or design control logic based on a single-scale (such as a single battery cell) electrochemical-thermal coupling model. While these methods can meet basic control requirements under simple operating conditions, they have significant limitations in complex dynamic environments: on the one hand, the dynamic coupling between multiple physical field parameters is not fully quantified, making it difficult to accurately model key characteristics such as thermal stress distribution caused by temperature gradients and battery life degradation; on the other hand, traditional control strategies lack the ability to dynamically perceive and adaptively adjust real-time operating states (such as road congestion, traffic flow changes, and vehicle load fluctuations), leading to a disconnect between control commands and actual needs, failing to meet the high-efficiency operation requirements of new energy heavy-duty trucks under multiple operating conditions.

[0003] In summary, existing technologies are insufficient to effectively optimize the high dynamics and multi-scale coupling effects of the battery power system of new energy heavy trucks, resulting in limited proactiveness, systematicity, and dynamic adaptability of control commands, and failing to meet the optimal operation requirements of the battery power system in complex scenarios. Summary of the Invention

[0004] This invention provides a method and system for optimizing the operation of a battery power system for new energy heavy-duty trucks. This method and system can accurately and efficiently optimize the operation of the battery power system for new energy heavy-duty trucks, effectively improving the system's energy efficiency, stability, and adaptability to different operating conditions.

[0005] An embodiment of the present invention provides a method for optimizing the operation of a battery power system for new energy heavy-duty trucks, comprising:

[0006] The current vehicle operating condition data of the new energy heavy truck is input into a pre-built decision network architecture. The state evaluation result is output through the state value function branch of the decision network architecture, and the action evaluation result is output through the action advantage function branch of the decision network architecture. The decision network architecture is built based on deep reinforcement learning.

[0007] The state evaluation results and action evaluation results are combined to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system.

[0008] A self-supervised learning method is used to extract features from multimodal operating data from different sensors of the battery power system of new energy heavy trucks to obtain a unified feature representation.

[0009] The unified feature representation is matched and evaluated with candidate control strategies, and the optimal control strategy is selected from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

[0010] As an improvement to the above solution, the current vehicle operating condition data of the new energy heavy truck is input into a pre-constructed decision network architecture. The state evaluation result is output through the state value function branch of the decision network architecture, and the action evaluation result is output through the action advantage function branch of the decision network architecture. The decision network architecture is constructed based on deep reinforcement learning and includes the following sub-steps:

[0011] The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture.

[0012] The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters.

[0013] The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data.

[0014] The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

[0015] As an improvement to the above scheme, the step of combining the state evaluation results with the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system includes the following sub-steps:

[0016] The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results.

[0017] The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result.

[0018] Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes.

[0019] Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

[0020] As an improvement to the above scheme, the step of using a self-supervised learning method to extract features from multimodal operating data from different sensors of the new energy heavy truck battery power system to obtain a unified feature representation includes the following sub-steps:

[0021] Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data;

[0022] The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data;

[0023] A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features.

[0024] The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs.

[0025] The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

[0026] As an improvement to the above scheme, the step of matching and evaluating the unified feature representation with candidate control strategies, and selecting the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system, includes the following sub-steps:

[0027] For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies.

[0028] The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values.

[0029] Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy.

[0030] The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

[0031] Another embodiment of the present invention provides an operation optimization system for a new energy heavy-duty truck battery power system, comprising:

[0032] The decision-making module is used to input the current vehicle operating condition data of the new energy heavy truck into a pre-built decision network architecture, output the state evaluation result through the state value function branch of the decision network architecture, and output the action evaluation result through the action advantage function branch of the decision network architecture; wherein, the decision network architecture is built based on deep reinforcement learning;

[0033] The strategy generation module is used to combine the state evaluation results and the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system.

[0034] The feature representation module is used to extract features from multimodal operating data from different sensors of the battery power system of new energy heavy trucks using a self-supervised learning method, and obtain a unified feature representation.

[0035] The optimization module is used to match and evaluate the unified feature representation with candidate control strategies, and select the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

[0036] As an improvement to the above solution, the decision-making module is specifically used for:

[0037] The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture.

[0038] The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters.

[0039] The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data.

[0040] The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

[0041] As an improvement to the above solution, the strategy generation module is specifically used for:

[0042] The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results.

[0043] The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result.

[0044] Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes.

[0045] Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

[0046] As an improvement to the above scheme, the feature representation module is specifically used for:

[0047] Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data;

[0048] The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data;

[0049] A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features.

[0050] The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs.

[0051] The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

[0052] As an improvement to the above solution, the optimization module is specifically used for:

[0053] For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies.

[0054] The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values.

[0055] Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy.

[0056] The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

[0057] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0058] First, a decision network architecture based on deep reinforcement learning is constructed. Current vehicle operating condition data is input into this architecture, and state evaluation results and action evaluation results are output through state value function branches and action advantage function branches, respectively. Leveraging the ability of deep reinforcement learning to autonomously learn the relationship between the environment and strategy, the architecture effectively evaluates the current state of the system and the potential value of different actions. Next, the two evaluation results are combined to generate multiple candidate control strategies, providing various options for system operation. Then, a self-supervised learning method is used to extract features from multimodal sensor operating data, obtaining a unified feature representation. Self-supervised learning can automatically mine data features, avoiding manual annotation and adapting to multi-source data, achieving a deep understanding of the battery power system's operating information. Finally, the unified feature representation is matched and evaluated with the candidate control strategies, and the optimal control strategy is selected for execution, ensuring that the strategy is adapted to the actual operating state of the system. In summary, compared to the problems of relying on fixed rules, simple data processing, and lack of precise decision-making in the background technology, this invention achieves dynamic adaptive decision-making through deep reinforcement learning and mines the value of multimodal data through self-supervised learning. This enables precise and efficient optimization of the operation of the new energy heavy-duty truck battery power system, effectively improving the system's energy efficiency, stability, and operating condition adaptability. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating an operation optimization method for a new energy heavy-duty truck battery power system according to an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the operation optimization system of a new energy heavy truck battery power system provided in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] See Figure 1 This is a flowchart illustrating an operation optimization method for a new energy heavy-duty truck battery power system according to an embodiment of the present invention. The operation optimization method for the new energy heavy-duty truck battery power system includes the following steps:

[0063] S10 inputs the current vehicle operating condition data of the new energy heavy truck into the pre-built decision network architecture, outputs the state evaluation result through the state value function branch of the decision network architecture, and outputs the action evaluation result through the action advantage function branch of the decision network architecture; wherein, the decision network architecture is built based on deep reinforcement learning;

[0064] S11, combine the state evaluation results with the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system;

[0065] S12, a self-supervised learning method is used to extract features from multimodal operation data from different sensors of the new energy heavy truck battery power system to obtain a unified feature representation;

[0066] S13, the unified feature representation is matched and evaluated with the candidate control strategies, and the optimal control strategy is selected from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

[0067] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0068] First, a decision network architecture based on deep reinforcement learning is constructed. Current vehicle operating condition data is input into this architecture, and state evaluation results and action evaluation results are output through state value function branches and action advantage function branches, respectively. Leveraging the ability of deep reinforcement learning to autonomously learn the relationship between the environment and strategy, the architecture effectively evaluates the current state of the system and the potential value of different actions. Next, the two evaluation results are combined to generate multiple candidate control strategies, providing various options for system operation. Then, a self-supervised learning method is used to extract features from multimodal sensor operating data, obtaining a unified feature representation. Self-supervised learning can automatically mine data features, avoiding manual annotation and adapting to multi-source data, achieving a deep understanding of the battery power system's operating information. Finally, the unified feature representation is matched and evaluated with the candidate control strategies, and the optimal control strategy is selected for execution, ensuring that the strategy is adapted to the actual operating state of the system. In summary, compared to the problems of relying on fixed rules, simple data processing, and lack of precise decision-making in the background technology, this invention achieves dynamic adaptive decision-making through deep reinforcement learning and mines the value of multimodal data through self-supervised learning. This enables precise and efficient optimization of the operation of the new energy heavy-duty truck battery power system, effectively improving the system's energy efficiency, stability, and operating condition adaptability.

[0069] As one example, the current vehicle operating condition data of the new energy heavy truck is input into a pre-constructed decision network architecture. The state evaluation result is output through the state value function branch of the decision network architecture, and the action evaluation result is output through the action advantage function branch of the decision network architecture. The decision network architecture is constructed based on deep reinforcement learning and includes the following sub-steps:

[0070] The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture.

[0071] The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters.

[0072] The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data.

[0073] The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

[0074] This embodiment constructs a deep reinforcement learning decision architecture based on the Actor-Critic network framework. Through multi-stage data processing and collaborative work with neural networks, it achieves accurate evaluation of the state of the battery power system of new energy heavy-duty trucks. First, a fully connected neural network is used to construct the state value function branch (Critic), which can capture the complex nonlinear relationships between high-dimensional operating condition data and output a value estimate of the system state. Simultaneously, a policy gradient network is used to construct the action advantage function branch (Actor), directly generating the optimal policy in the continuous action space. This dual-branch architecture retains the stability of the value function method while possessing the flexibility of the policy gradient method. Second, feature engineering is used to process the current vehicle operating condition data, extracting key parameters such as battery SOC, temperature, and vehicle load, and standardizing them to eliminate the influence of different dimensions on network training and ensure the standardization of input data. Finally, the preprocessed data is simultaneously input into both branches: the state value function branch outputs the value assessment of the current system state, and the action advantage function branch outputs the degree of advantage of different control actions, forming state-action value pairs. The advantages of this technical solution are as follows: On the one hand, the dual-branch structure enables efficient exploration and utilization of the continuous action space, solving the problem of convergence difficulty in high-dimensional complex systems in traditional reinforcement learning; on the other hand, feature engineering improves the network's adaptability to multi-source heterogeneous data, enhances the model's generalization ability under different operating conditions, and ultimately achieves accurate quantitative evaluation of the operating status of the battery power system, providing a reliable basis for subsequent control strategy optimization.

[0075] Specifically, the working process of this embodiment is as follows:

[0076] First, the decision network architecture is initialized. This embodiment uses the Actor-Critic network framework to construct the decision network architecture, which consists of two key branches: the state value function branch and the action advantage function branch. The state value function branch employs a fully connected neural network structure, and its function is to evaluate the value of the system's current state. The fully connected neural network is interconnected through multiple neuron layers. Each neuron receives the outputs of all neurons in the previous layer and processes the input data through an activation function, progressively extracting high-level features from the data. In this embodiment, the fully connected neural network contains three hidden layers, each with 64, 32, and 16 neurons respectively. By continuously adjusting the connection weights between neurons, an accurate evaluation of the system state value is achieved.

[0077] The action advantage function branch employs a policy gradient network structure to evaluate the advantage of different actions in the current state. The policy gradient network, based on the policy gradient algorithm, maximizes long-term cumulative reward by optimizing the policy function. During initialization, the learning rate of the policy gradient network is set to 0.001, and the discount factor is set to 0.99. The initial decision network architecture is obtained by randomly initializing the parameters in the network.

[0078] Next, the current operating condition data of the new energy heavy-duty truck is acquired. Various sensors are deployed on the vehicle for real-time data collection. The battery management system collects the battery's current state of charge (SOC) through built-in voltage and current sensors, and battery temperature through temperature sensors; vehicle speed sensors are installed on the wheels or transmission system to obtain the vehicle's speed in real time; vehicle load information is measured by pressure sensors installed on the suspension or axles; and road gradient information is calculated using the onboard inertial measurement unit (IMU) combined with GPS data. After preliminary processing by the onboard controller, the data collected by these sensors is aggregated to form current vehicle operating condition data containing various parameters.

[0079] Then, feature engineering is performed on the acquired current vehicle operating condition data. Since the raw data may contain noise and redundant information, and the numerical ranges and units of the data vary, it cannot be directly input into the decision network architecture. Therefore, feature extraction and standardization are required. For parameters such as the battery's current state of charge, battery temperature, vehicle speed, vehicle load, and road slope, obvious outliers are first removed using statistical analysis methods. Then, a normalization method is used to map the data to the [0,1] interval. The specific normalization formula is as follows: Where, x normHere are the normalized data, and x is the original data. max and x min These represent the maximum and minimum values ​​of the parameter in the historical data. After normalization, the data meets the input requirements of the initialized decision network architecture, resulting in preprocessed current vehicle operating condition data.

[0080] Finally, the preprocessed current vehicle operating condition data is input into the initialized decision network architecture. After receiving the input data, the state value function branch performs calculations sequentially through the multi-layer neurons of the fully connected neural network. Each neuron, based on the input data and connection weights, is processed by an activation function (ReLU activation function is used in this embodiment) to gradually extract key features from the data, and finally outputs a scalar value. This scalar value reflects the value of the current system state, i.e., the state evaluation result.

[0081] When processing the input data, the action advantage function branch, based on the policy gradient algorithm, calculates the advantage of different actions in the current state by continuously adjusting the parameters of the policy function. The policy function can be represented as π(a|s; θ), where a represents the action, s represents the state, and θ represents the parameters of the policy function. The gradient of the policy function is then calculated. The parameters θ are updated according to the gradient direction to increase the probability of a better action in the current state. The final output is the action evaluation result representing the degree of advantage of different actions in the current state.

[0082] As one example, combining the state evaluation results with the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system includes the following sub-steps:

[0083] The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results.

[0084] The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result.

[0085] Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes.

[0086] Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

[0087] In this embodiment, by combining multi-stage data processing and algorithms, candidate control strategies that meet the operational requirements of the battery power system of new energy heavy-duty trucks are generated based on the state and action evaluation results. First, the state and action evaluation results are dimensionally matched to resolve data structure differences and ensure effective fusion, laying the foundation for subsequent strategy generation. Next, the weights are dynamically adjusted according to the importance of states and actions under different operating conditions, and the dimensionally matched evaluation results are weighted and fused to achieve a comprehensive consideration of the value of system states and actions, making the fused evaluation results more consistent with actual operational needs. Then, the Monte Carlo tree search algorithm is used to search the action space. This algorithm efficiently explores the complex action space through random sampling and simulation, generating multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, rapidly expanding the range of strategy selection. Finally, the feasibility of the preliminary control scheme is verified by checking constraints such as battery charging and discharging limits, temperature limits, and vehicle power requirements, selecting multiple candidate control strategies that meet actual operating conditions, and avoiding the generation of infeasible control schemes. In summary, this embodiment, through scientific data processing and algorithm application, realizes the transformation from evaluation results to feasible candidate control strategies, effectively improving the comprehensiveness, rationality, and feasibility of control strategy generation, avoiding blind decision-making, providing diversified and feasible strategy options for the operation optimization of the battery power system of new energy heavy trucks, enhancing the stability and efficiency of the system under different operating conditions, and enabling it to better adapt to complex and ever-changing operating environments.

[0088] Specifically, the working process of this embodiment is as follows:

[0089] First, the state evaluation results and action evaluation results undergo dimensionality matching. The state evaluation result is a scalar output from the state value function branch of the decision network architecture, used to assess the value of the current system state. The action evaluation result is a vector output from the action advantage function branch, where each element corresponds to the degree of advantage of a pre-defined battery power system adjustment action (such as adjusting charging power, discharging power, or initiating battery cooling) in the current state. Since the two have different dimensions, they cannot be directly combined. Therefore, the scalar state evaluation result needs to be expanded into a vector with the same dimension as the action evaluation result. Specifically, based on the vector length n of the action evaluation result, the state evaluation result is copied n times to form a new vector, making the dimensions consistent, thus obtaining the dimensionally matched state evaluation result and action evaluation result.

[0090] Next, weighted fusion is performed. To achieve a dynamic trade-off between the importance of states and actions under different operating conditions, this embodiment improves the traditional weighting method. A dynamic weight generation model based on operating condition parameters is constructed. This model takes real-time collected operating condition data of new energy heavy trucks (including battery state of charge, temperature, vehicle speed, load, road slope, etc.) as input, and outputs the state evaluation result weight w through a pre-trained multilayer perceptron network. s Weight w of action evaluation results a And satisfy w s +w a =1. For example, when the battery state of charge is below 20%, the model will increase w. s The value of w is increased to make the system pay more attention to the battery status to ensure safety; when the vehicle is under heavy load and climbing conditions, w is increased. a The value of [value] is prioritized to meet the power output requirements. The state evaluation result vector S and the action evaluation result vector A, after dimension matching, are then processed according to the formula R = w s ×S+w a The weighted fusion of ×A yields the fused evaluation result vector R, which comprehensively reflects the combined value of the current state and actions.

[0091] Based on the fused evaluation result R, an improved Monte Carlo tree search algorithm is used to search the action space. Traditional Monte Carlo tree search relies solely on the upper confidence bound (UCB) formula to evaluate node value during selection. This embodiment introduces a heuristic evaluation factor H based on the working condition. The improved node selection formula is as follows: Where Q(s, a) represents the cumulative average reward for performing action a in state s, N(s) is the total number of times state s is visited, N(s, a) is the number of times action a is selected in state s, c is the exploration coefficient used to balance exploration and utilization, and H(s, a) is a heuristic evaluation value calculated based on the current operating conditions (such as battery status, vehicle load, etc.), aiming to guide the search in a direction that better meets the needs of the current operating conditions. During the search process, starting from the root node (current system state), based on UCB... new The formula selects child nodes layer by layer. When an expandable node is reached, all possible actions are added as child nodes. Simulation then begins from the new node. During the simulation, termination conditions are set based on the current operating conditions (such as the battery reaching its limit or the vehicle completing a certain distance). The cumulative reward value of the simulation is recorded and propagated back along the search tree to update the statistics of all nodes on the path. Through more than 1000 iterations, multiple possible action sequences are generated, each corresponding to a control scheme for the battery power system, forming a preliminary set of control schemes.

[0092] Finally, the feasibility of several preliminary control schemes was verified. Regarding battery charge / discharge limitations, the minimum charging power P was determined based on parameters provided by the battery manufacturer. min and maximum discharge power P max Check whether the battery charging and discharging power P in each scheme meets the P requirement. min ≤P≤P max Regarding battery temperature limits, real-time monitoring data is used from temperature sensors on the vehicle, combined with a battery thermal management model to predict the battery temperature T after the implementation of the scheme, to determine whether it is within the safe temperature range T. min ≤T≤T max Regarding vehicle power demand constraints, the required power P is calculated using a vehicle dynamics model based on parameters such as the vehicle's current speed, load, and road gradient. req To ensure the power provided by the solution P offer Satisfy P offer ≥P req Only preliminary control schemes that simultaneously meet the above three constraints will be selected as multiple candidate control strategies that meet the requirements, and used for subsequent decision-making in optimizing the operation of the battery power system of new energy heavy trucks.

[0093] As one example, the method of using self-supervised learning to extract features from multimodal operating data from different sensors of the new energy heavy-duty truck battery power system to obtain a unified feature representation includes the following sub-steps:

[0094] Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data;

[0095] The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data;

[0096] A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features.

[0097] The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs.

[0098] The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

[0099] In this embodiment, a self-supervised learning method is used to perform deep processing on the multi-source heterogeneous sensor data of the new energy heavy-duty truck battery power system, achieving efficient feature extraction and unified representation. First, raw multimodal operating data from different sensors such as voltage, current, and temperature are collected to obtain basic information about the system's operation. Then, preprocessing is performed, including data cleaning and missing value imputation, to improve data quality. Next, a self-supervised learning model based on a contrastive learning framework is constructed, where the encoder network is responsible for extracting key features from the data, and the decoder network verifies the effectiveness of the features through data reconstruction. During training, by maximizing the similarity of positive sample pairs (similar data) and minimizing the similarity of negative sample pairs (different data), the model is guided to autonomously learn the inherent patterns of the data, avoiding the limitations of traditional supervised learning that relies on a large amount of labeled data. Finally, the trained encoder network is used to extract features from the preprocessed data, mapping the multimodal data to a unified feature space, eliminating modal differences, and forming a unified feature representation that can be used for subsequent decision-making. In summary, this embodiment achieves unsupervised feature extraction from multimodal data through self-supervised learning, reducing data annotation costs and reliance on human labor. Simultaneously, it effectively mines deep features of the data, improving the accuracy and generalization of feature representation. The unified feature representation provides standardized input for subsequent matching and evaluation of candidate control strategies, ensuring that strategy selection accurately matches the actual operating state of the system. This effectively improves the efficiency and accuracy of operational optimization decisions for the battery power system of new energy heavy-duty trucks, enhancing the system's adaptability to complex operating conditions.

[0100] Specifically, the working process of this embodiment is as follows:

[0101] When collecting multimodal operational data from various sensors in the battery power system of new energy heavy-duty trucks, the sensors on the vehicle perform their respective functions and work together. Various sensors within the battery management system monitor the battery status in real time. Voltage sensors accurately measure the voltage of individual battery cells and the overall battery structure to reflect the battery's electrical characteristics; current sensors continuously record the magnitude of charging and discharging current, providing a basis for calculating key parameters such as the battery's state of charge (SOC); temperature sensors are distributed at various key locations within the battery module, including cell surfaces and connection points, collecting battery temperature information at high frequency to prevent abnormal temperatures from affecting battery performance and lifespan. Furthermore, vehicle speed sensors obtain real-time vehicle speed by sensing wheel rotation or driveshaft rotation; pressure sensors installed on the axle, suspension, and other parts accurately sense the vehicle's load weight; and the inertial measurement unit (IMU), combined with the global positioning system (GPS), can not only obtain information such as vehicle acceleration and attitude angles but also accurately calculate the gradient of the road the vehicle is traveling on. These sensors continuously collect data at their respective set frequencies. For example, the temperature sensor collects data 10 times per second, and the GPS sensor collects data once per second. The collected data is packaged and transmitted through the vehicle controller's local area network according to a specific data protocol format, and finally aggregated to form raw multimodal operating data containing various types and different timestamps.

[0102] When preprocessing the raw multimodal operational data, the first challenge is the inconsistent data sampling frequencies. To address this, the sensor with the highest sampling frequency (assumed to be a temperature sensor, 10Hz) is used as the benchmark. For other low-frequency data (such as 1Hz GPS data), a cubic spline interpolation algorithm is employed. This algorithm constructs a smooth curve function based on known data points and inserts appropriate data values ​​between adjacent data points, ensuring precise alignment of all data on the time axis. Next, dimensional differences are addressed. For numerical data such as voltage (unit: volt) and current (unit: ampere), an improved normalization method is used. First, the mean μ and standard deviation σ of each sensor's data over historical operations are calculated, and then the formula is used to... Normalization is performed, where x is the original data and ∈ is a very small constant to prevent calculation errors when the standard deviation is 0. The normalized data is mapped to the interval [-1, 1]. For categorical data, such as sensor fault labels (normal, minor fault, severe fault), an extended one-hot encoding method is used. In addition to the conventional binary encoding, fault probability weights are added, making the encoding more reflective of the actual fault situation. Finally, an adaptive median filtering algorithm is used to remove noise from the data. This algorithm dynamically adjusts the size of the filtering window based on the statistical characteristics of local data regions. Compared with fixed-window filtering, it can better preserve data details, thus obtaining high-quality preprocessed multimodal operating data.

[0103] When constructing the self-supervised learning model, a novel dual-path contrastive learning framework is adopted. The encoder network consists of parallel feature extraction paths. One path uses an improved multi-scale convolutional neural network, containing convolutional layers with different kernel sizes (3×3, 5×5, 7×7). Through convolutional operations at different scales, features of different granularities in multimodal data can be captured. The other path adopts a structure combining a Long Short-Term Memory (LSTM) network with an attention mechanism. LSTM can effectively handle long-term dependencies in time-series data, while the attention mechanism can focus on key information in the data. The features extracted by the two paths are fused in the fusion layer through weighted summation. The weights are dynamically generated by a small fully connected neural network based on the data features. The decoder network is designed using the idea of ​​a Generative Adversarial Network (GAN), including a generator and a discriminator. The generator is responsible for reconstructing the feature vector output by the encoder into the original data form, while the discriminator is used to judge the authenticity of the reconstructed data and the original data. Through adversarial training between the two, the reconstruction capability of the decoder is improved. In addition, a modal cross-regularization term is introduced into the model to constrain the distribution of different modal data in the feature space, making it more consistent. Its calculation formula is as follows: Where μ i and μ j These are the mean vectors of the i-th and j-th modal data features, respectively, ∑ i and ∑ j These are the corresponding covariance matrices, ||·||2 is the L2 norm, and ||·|| F It is the Frobenius norm.

[0104] When preprocessed multimodal runtime data is input into the self-supervised learning model for training, a contrastive learning strategy with dynamic difficulty adjustment is employed. Regarding sample pair generation, in addition to considering temporal correlation and modal similarity, the concept of data entropy is introduced. For data samples with high entropy values ​​(large data fluctuations and high uncertainty), their probability of being generated as positive sample pairs is increased, as such data often contains more useful information. In the loss function calculation, the traditional InfoNCE loss function is improved by introducing a dynamic temperature parameter τ(t), whose formula varies with the training epoch t is as follows: Where τ0 is the initial temperature, T is the total number of training epochs, and α is the adjustment coefficient. In the early stages of training, a larger τ(t) value makes the model more focused on the alignment of the overall feature distribution; as training progresses, τ(t) gradually decreases, making the model more sensitive to subtle differences between samples. During training, a multi-stage optimization strategy is also employed: first, the two paths of the encoder and the decoder are trained separately, enabling each part to initially possess good feature extraction and reconstruction capabilities; then, overall joint training is performed, minimizing the total loss function L = L... contrast +λ1L recon +λ2Lcross L contrast It is a contrastive learning loss, L recon It is the reconstruction loss, L cross λ1 and λ2 are the modality cross-regularization loss and the balance coefficients. After 100 training rounds, a well-trained self-supervised learning model is obtained.

[0105] When using a pre-trained self-supervised learning model to extract features from preprocessed multimodal runtime data, the data is input in batches. Inside the encoder, a multi-scale convolutional neural network path first performs convolution operations at different scales on the data. Small convolutional kernels (3×3) capture local detailed features, while large convolutional kernels (7×7) acquire global structural features. After multiple convolutional and pooling operations, a preliminary feature map is output. The LSTM-attention mechanism path processes the data sequentially in chronological order. LSTM units remember long-term dependencies, and the attention mechanism dynamically adjusts the level of attention to historical information based on the current input data, outputting a vector containing time-series features. The outputs of the two paths are weighted and fused in a fusion layer based on weights calculated by a fully connected neural network to obtain a fused feature representation. Finally, a global average pooling layer compresses the fused features into a fixed-dimensional (e.g., 512-dimensional) feature vector, successfully mapping data from different modalities to a unified feature space, resulting in a unified feature representation with rich semantic information and high representativeness.

[0106] As one example, the process of matching and evaluating the unified feature representation with candidate control strategies, and selecting the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy-duty truck battery power system, includes the following sub-steps:

[0107] For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies.

[0108] The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values.

[0109] Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy.

[0110] The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

[0111] In this embodiment, a correlation mechanism is established between a unified feature representation and candidate control strategies. Through quantitative matching and screening, precise decision-making for optimizing the operation of the battery power system of new energy heavy-duty trucks is achieved. First, based on the action sequence and expected system response in the candidate control strategies, a corresponding feature vector representation is constructed for each candidate control strategy, transforming the abstract control strategy into computable feature data and establishing a digital expression of the strategy. Next, the cosine similarity measurement method is used to calculate the similarity between the unified feature representation and the feature vectors of each candidate control strategy, using quantitative indicators to measure the degree of fit between the strategy and the current operating state of the system. Then, the candidate control strategies are ranked according to the calculated similarity values, and the strategy with the highest similarity value is selected as the optimal control strategy, ensuring that the selected strategy has the best match with the actual state of the system. Finally, the optimal control strategy is transformed into specific control commands, such as battery charging and discharging current adjustment and motor power output control, realizing the transformation from strategy selection to actual control. In summary, this embodiment, through a scientifically quantitative matching and evaluation mechanism, avoids the blindness and subjectivity of control strategy selection, and can quickly and accurately select the optimal solution that best suits the current operating state of the system from multiple candidate control strategies. This precise decision-making mechanism effectively improves the efficiency and accuracy of optimizing the operation of the battery power system of new energy heavy trucks, ensuring that the system always operates in the optimal state under complex and ever-changing working conditions, effectively improving battery life, reducing energy consumption, and enhancing the overall stability and reliability of the system.

[0112] Specifically, the working process of this embodiment is as follows:

[0113] In practical applications, a corresponding feature vector representation is first generated for each candidate control strategy. Each candidate control strategy is essentially a sequence of multiple control actions; for example, in a battery power system, it might include actions such as adjusting charging power, discharging current, and activating the thermal management system. To construct the feature vector, this embodiment innovatively introduces a time-weighted encoding mechanism. For each action, not only is the action type itself considered (e.g., "increase charging power by 20kW"), but also the execution time, expected duration, and expected system response (e.g., battery SOC change rate, temperature change trend, etc.) are taken into account. Specifically, a control strategy is decomposed into multiple time windows, and the actions and expected responses within each window are encoded as high-dimensional vectors. For example, within the t-th time window, the action type is represented by vector A using one-hot encoding. t The expected rate of change of SOC and the rate of change of temperature are expressed as scalars ΔSOC. t and ΔT t They are concatenated and subjected to nonlinear transformation to obtain the feature vector V of this time window. t =σ(W1·[A t ΔSOC t ;ΔTt ]+b1), where σ is the activation function, and W1 and b1 are learnable parameters. Then, the feature vectors of all time windows are summed in a time-weighted manner, with the weights determined by a Gaussian distribution function, where windows closer to the current time have larger weights. This process ultimately yields the feature vector representation of the candidate control policy. in T is the total number of time windows, and σ is the adjustment parameter.

[0114] Next, the similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. This embodiment improves upon traditional cosine similarity by proposing an adaptive cosine similarity measurement method. Traditional cosine similarity only considers the consistency of vector direction, while this embodiment also considers the magnitude information of the vectors, because the magnitude reflects the importance of the feature to a certain extent. The formula for the adaptive cosine similarity measurement in this embodiment is as follows: Where u is the unified feature representation, v is the feature vector of the candidate control strategy, and λ is the adjustment parameter. When the magnitudes of the two vectors differ significantly, the second term will appropriately reduce the similarity value to avoid misjudgment caused by the difference in magnitudes. The similarity value between the unified feature representation and the feature vector of each candidate control strategy is calculated using this formula, forming a similarity set.

[0115] Then, the candidate control strategies are ranked based on multiple similarity values. This embodiment employs a constrained ranking algorithm that considers not only the magnitude of the similarity values ​​but also the feasibility and stability of the control strategies. Specifically, a comprehensive scoring function Score(s) is defined. i ) = Sim adapt (u, v) i )+β1·F(s i )+β2·S(s i ), where s i Let Sim be the i-th candidate control policy. adapt (u, v) i F(s) is its adaptive cosine similarity to the unified feature representation. i S(s) is the feasibility score (calculated based on constraints such as battery charge / discharge limits and temperature limits). i β1 and β2 are the stability score (obtained by evaluating the volatility of the policy within a certain time window) and the balance coefficients, respectively. Candidate control policies are ranked in descending order based on this comprehensive score, and the control policy with the highest score is selected as the optimal control policy. This method avoids the problem of selecting unstable or infeasible policies that may result from simply ranking based on similarity.

[0116] Finally, the optimal control strategy is transformed into specific control commands. This embodiment designs a hierarchical command generation mechanism to transform the abstract control strategy into executable hardware commands. First, based on the action sequence in the optimal control strategy, the battery system components that need to be adjusted (such as chargers, inverters, cooling systems, etc.) are determined. Then, for each component, an action-command mapping table is established; for example, "increase charging power by 20kW" corresponds to the specific control parameter adjustment of the charger. Considering the inertia and response delay of the actual system, predictive control is adopted to optimize the control commands in a time series. By establishing a dynamic model of the battery system, the system state after each control command is executed is predicted, and the timing and intensity of command execution are adjusted to ensure a smooth transition of the system to the expected state. The finally generated control commands are sent to each actuator through the vehicle's Controller Area Network (CAN) bus to achieve operational optimization of the new energy heavy-duty truck battery power system.

[0117] See Figure 2 This is a schematic diagram of the operation optimization system for a new energy heavy-duty truck battery power system according to an embodiment of the present invention. The operation optimization system for the new energy heavy-duty truck battery power system includes:

[0118] The decision module 10 is used to input the current vehicle operating condition data of the new energy heavy truck into the pre-built decision network architecture, output the state evaluation result through the state value function branch of the decision network architecture, and output the action evaluation result through the action advantage function branch of the decision network architecture; wherein, the decision network architecture is built based on deep reinforcement learning;

[0119] The strategy generation module 11 is used to combine the state evaluation results and the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system.

[0120] Feature representation module 12 is used to extract features from multimodal operating data from different sensors of the new energy heavy truck battery power system using a self-supervised learning method to obtain a unified feature representation;

[0121] The optimization module 13 is used to match and evaluate the unified feature representation with the candidate control strategies, and select the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

[0122] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0123] First, a decision network architecture based on deep reinforcement learning is constructed. Current vehicle operating condition data is input into this architecture, and state evaluation results and action evaluation results are output through state value function branches and action advantage function branches, respectively. Leveraging the ability of deep reinforcement learning to autonomously learn the relationship between the environment and strategy, the architecture effectively evaluates the current state of the system and the potential value of different actions. Next, the two evaluation results are combined to generate multiple candidate control strategies, providing various options for system operation. Then, a self-supervised learning method is used to extract features from multimodal sensor operating data, obtaining a unified feature representation. Self-supervised learning can automatically mine data features, avoiding manual annotation and adapting to multi-source data, achieving a deep understanding of the battery power system's operating information. Finally, the unified feature representation is matched and evaluated with the candidate control strategies, and the optimal control strategy is selected for execution, ensuring that the strategy is adapted to the actual operating state of the system. In summary, compared to the problems of relying on fixed rules, simple data processing, and lack of precise decision-making in the background technology, this invention achieves dynamic adaptive decision-making through deep reinforcement learning and mines the value of multimodal data through self-supervised learning. This enables precise and efficient optimization of the operation of the new energy heavy-duty truck battery power system, effectively improving the system's energy efficiency, stability, and operating condition adaptability.

[0124] As one example, the decision module is specifically used for:

[0125] The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture.

[0126] The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters.

[0127] The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data.

[0128] The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

[0129] As one example, the policy generation module is specifically used for:

[0130] The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results.

[0131] The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result.

[0132] Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes.

[0133] Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

[0134] As one example, the feature representation module is specifically used for:

[0135] Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data;

[0136] The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data;

[0137] A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features.

[0138] The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs.

[0139] The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

[0140] As one example, the optimization module is specifically used for:

[0141] For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies.

[0142] The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values.

[0143] Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy.

[0144] The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

[0145] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0146] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for optimizing the operation of a battery power system for new energy heavy-duty trucks, characterized in that, include: The current vehicle operating condition data of the new energy heavy truck is input into a pre-built decision network architecture. The state evaluation result is output through the state value function branch of the decision network architecture, and the action evaluation result is output through the action advantage function branch of the decision network architecture. The decision network architecture is built based on deep reinforcement learning. The state evaluation results and action evaluation results are combined to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system. A self-supervised learning method is used to extract features from multimodal operating data from different sensors of the battery power system of new energy heavy trucks to obtain a unified feature representation. The unified feature representation is matched and evaluated with candidate control strategies, and the optimal control strategy is selected from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

2. The operation optimization method for the battery power system of new energy heavy trucks as described in claim 1, characterized in that, The process involves inputting the current vehicle operating condition data of the new energy heavy-duty truck into a pre-constructed decision network architecture, outputting state evaluation results through the state value function branch of the decision network architecture, and outputting action evaluation results through the action advantage function branch of the decision network architecture. The decision network architecture is built based on deep reinforcement learning and includes the following sub-steps: The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture. The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters. The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data. The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

3. The operation optimization method for the battery power system of new energy heavy trucks as described in claim 1, characterized in that, The step of combining the state evaluation results with the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system includes the following sub-steps: The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results. The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result. Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes. Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

4. The operation optimization method for the battery power system of new energy heavy trucks as described in claim 1, characterized in that, The step of using a self-supervised learning method to extract features from multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain a unified feature representation includes the following sub-steps: Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data; The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data; A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features. The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs. The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

5. The operation optimization method for the battery power system of new energy heavy trucks as described in claim 1, characterized in that, The process of matching and evaluating the unified feature representation with candidate control strategies, and selecting the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy-duty truck battery power system, includes the following sub-steps: For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies. The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values. Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy. The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

6. An operation optimization system for a new energy heavy-duty truck battery power system, characterized in that, include: The decision-making module is used to input the current vehicle operating condition data of the new energy heavy truck into a pre-built decision network architecture, output the state evaluation result through the state value function branch of the decision network architecture, and output the action evaluation result through the action advantage function branch of the decision network architecture; wherein, the decision network architecture is built based on deep reinforcement learning; The strategy generation module is used to combine the state evaluation results and the action evaluation results to generate multiple candidate control strategies for optimizing the operation of the new energy heavy truck battery power system. The feature representation module is used to extract features from multimodal operating data from different sensors of the battery power system of new energy heavy trucks using a self-supervised learning method, and obtain a unified feature representation. The optimization module is used to match and evaluate the unified feature representation with candidate control strategies, and select the optimal control strategy from multiple candidate control strategies to optimize the operation of the new energy heavy truck battery power system.

7. The operation optimization system for the battery power system of new energy heavy trucks as described in claim 6, characterized in that, The decision-making module is specifically used for: The pre-built decision network architecture is initialized. This decision network architecture adopts the Actor-Critic network framework, the state value function branch adopts a fully connected neural network structure, and the action advantage function branch adopts a policy gradient network structure, thus obtaining the initialized decision network architecture. The system acquires current vehicle operating condition data for new energy heavy trucks, including information on the current state of charge of the battery, battery temperature, vehicle speed, vehicle load, and road gradient, to obtain current vehicle operating condition data containing multiple parameters. The current vehicle operating condition data is processed by feature engineering, key features are extracted and standardized to make the data meet the input requirements of the initialized decision network architecture, thus obtaining the preprocessed current vehicle operating condition data. The preprocessed current vehicle operating condition data is input into the initialized decision network architecture. The state value function branch processes the input data and outputs a state evaluation result that reflects the value of the current system state. The action advantage function branch processes the input data and outputs an action evaluation result that represents the degree of advantage of different actions in the current state.

8. The operation optimization system for the battery power system of new energy heavy trucks as described in claim 6, characterized in that, The strategy generation module is specifically used for: The state assessment results and action assessment results are subjected to dimensional matching to ensure that they can be effectively combined in terms of dimensions, resulting in dimensionally matched state assessment results and action assessment results. The state evaluation results and action evaluation results after dimension matching are weighted and fused. The weights are dynamically adjusted according to the importance of state and action under different working conditions to obtain the fused evaluation result. Based on the fused evaluation results, the Monte Carlo tree search algorithm is used to search the action space to generate multiple possible action sequences. Each action sequence corresponds to a control scheme for the battery power system, resulting in multiple preliminary control schemes. Feasibility verification was conducted on several preliminary control schemes to check whether each scheme met the constraints of battery charge / discharge limits, temperature limits, and vehicle power requirements, and to select several candidate control strategies that met the constraints.

9. The operation optimization system for the battery power system of new energy heavy trucks as described in claim 6, characterized in that, The feature representation module is specifically used for: Collect multimodal operating data from different sensors in the battery power system of new energy heavy trucks to obtain raw multimodal operating data; The original multimodal operation data is preprocessed to obtain preprocessed multimodal operation data; A self-supervised learning model is constructed, which adopts a contrastive learning framework and includes an encoder network and a decoder network. The encoder network is used to extract data features, and the decoder network is used to reconstruct data from the features. The preprocessed multimodal running data is input into the self-supervised learning model. The model is trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs. The encoder network, which uses a pre-trained self-supervised learning model, extracts features from the pre-processed multimodal running data, mapping the data from different modalities to a unified feature space to obtain a unified feature representation.

10. The operation optimization system for the battery power system of new energy heavy trucks as described in claim 6, characterized in that, The optimization module is specifically used for: For each candidate control strategy, a corresponding feature vector representation is generated. This feature vector representation is constructed based on the action sequence in the candidate control strategy and the expected system response, resulting in feature vectors for multiple candidate control strategies. The similarity between the unified feature representation and the feature vector of each candidate control strategy is calculated. The similarity calculation adopts the cosine similarity measurement method to obtain multiple similarity values. Candidate control strategies are ranked according to multiple similarity values, and the candidate control strategy with the highest similarity value is selected as the optimal control strategy. The optimal control strategy is translated into specific control commands, and the operation of the battery power system of new energy heavy trucks is optimized by executing these control commands.

Images