Vehicle level balancing method and system based on spatio-temporal causal graph convolution and intent load

Abstract

The application discloses a whole vehicle power balance method and system based on spatiotemporal causal graph convolution and intention load cooperation. The method extracts pseudo-synchronous input features from a multi-modal buffer to construct a dynamic graph, and uses a dynamic edge weight generator and a spatiotemporal causal graph convolution layer to extract global graph features. The input intention perception predictor is parallelly outputted with a continuous intention encoding vector and a future load power sequence through a causal mask transformer encoder and a conditional time series convolution network, and semantic alignment is realized through intention-load consistency loss. An enhanced observation state containing an intention vector and a load prediction is constructed, and multi-agent reinforcement learning is used for joint training to generate a control strategy by weighting atomic rewards according to intention intensity. The application realizes multi-modal spatiotemporal feature alignment and intention-load cooperative decision-making, and significantly improves the energy efficiency and dynamic response accuracy of whole vehicle power balance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for vehicle electrical systems, specifically relating to a vehicle electrical balancing method and system based on spatiotemporal causal graph convolution and intentional load coordination. Background Technology

[0002] As vehicles become increasingly electrified and intelligent, the load on the vehicle's low-voltage electrical system becomes more complex. High-power loads such as advanced driver assistance systems, infotainment systems, and multi-zone climate control place higher demands on the dynamic response capability of electrical balance. Traditional rule-based or PID-based energy management strategies mostly focus on maintaining voltage stability. When faced with sudden load changes and driver intention shifts, they can often only respond passively and are unable to achieve proactive adjustment.

[0003] In recent years, academia and industry have begun to explore the use of artificial intelligence to enhance the cognitive and decision-making capabilities of electric balancing systems. While some research has attempted to incorporate multimodal fusion, graph neural networks, or multi-agent reinforcement learning to improve the intention recognition and energy management capabilities of electric balancing systems, existing methods generally suffer from a disconnect between "intention understanding" and "load control": multimodal fusion often uses fixed weights, failing to adapt to dynamic environmental changes; there is a lack of feedback loops between intention recognition and energy management; and multi-agent optimization objectives are fixed, making it difficult to handle continuous changes and coexistence of intentions. The common problem is that while the system can recognize driver intentions, it cannot establish causal consistency constraints between intentions and future loads, causing electric balancing management to remain at a passive response level.

[0004] Therefore, there is an urgent need for a vehicle electric balance method and system that can establish a causal reasoning link from intent semantics to load change, in order to solve the technical problem of the separation between intent understanding and load control in existing vehicle electric balance management technologies. Summary of the Invention

[0005] This invention aims to solve the technical problem of the separation between intent understanding and load control in existing vehicle electric balance management technologies, and provides a vehicle electric balance method and system that can establish a causal reasoning link from intent semantics to load change.

[0006] To achieve the above objectives, the present invention provides a vehicle electric balance method that combines spatiotemporal causal graph convolution with intentional load coordination, comprising the following steps: Step 1: Extract the feature vector with the timestamp closest to the current moment from the feature buffer of each modality at a fixed period, and use it as the pseudo-synchronization input feature of each modality; Step 2: Construct dynamic graph nodes from the pseudo-synchronous input features of each modality, and generate a dynamic soft adjacency matrix through a dynamic edge weight generator; construct a spatiotemporal causal graph convolutional layer, and use the spatiotemporal causal graph convolutional layer to perform spatial aggregation and time propagation of gated recurrent units on the dynamic soft adjacency matrix and the current node features. After stacking multiple layers, obtain global graph features and dominant modality identifiers through multi-head graph attention pooling; cache the global graph features of each historical moment in chronological order as a global graph feature sequence; Step 3: Construct an intent-aware spatiotemporal state predictor. Input the global graph feature sequence into the spatiotemporal state predictor, which encodes the intent state sequence and outputs the intent classification probability and continuous intent encoding vector. Using the intent state sequence and continuous intent encoding vector as conditions, the conditional temporal convolutional prediction network of the spatiotemporal state predictor outputs the future load power sequence. During the training phase, a total loss function including intent-load consistency loss is used for optimization. Step 4: Each control agent collects its own physical state vector; constructs the enhanced observation state of each control agent, which includes the physical state vector, dominant mode identifier, continuous intent encoding vector and future load power sequence, for joint training of multi-agent deep reinforcement learning; after joint training, the vehicle outputs control actions according to the current enhanced observation state to realize the vehicle's electric balance management.

[0007] Furthermore, the dynamic edge weight generator mentioned in step 2 is a multilayer perceptron. The input of the dynamic edge weight generator is the concatenation of the features of all current nodes and the global graph features from the previous time step, and the output is a row-normalized dynamic soft adjacency matrix. Dynamic edge weight generator: ; ; in, The global graph features representing the previous decision-making moment; Represents a dynamic edge weight generator network; This represents performing Softmax normalization on each row of the matrix; This represents the generated dynamic soft adjacency matrix. For matrix elements at time... From the source node mode to target node mode The weight of information transmission strength or credibility; These represent the feature representations of four modalities at the initial moment: visual, speech and semantic, vehicle dynamics, and environment. To output a normalized dynamic soft adjacency matrix.

[0008] Furthermore, the spatiotemporal causal graph convolutional layer described in step 2 includes a graph convolution module and a gated recurrent unit. Spatial aggregation is based on the weighted fusion of neighbor node features using a dynamic soft adjacency matrix and mapped by an activation function. Temporal propagation uses the gated recurrent unit to fuse the current aggregated features with the hidden state of the previous time step.

[0009] Furthermore, in step 2, the multi-head graph attention pooling, while outputting global graph features, determines the dominant mode identifier at the current moment by summing the attention weights of each mode and taking the mode corresponding to the maximum value.

[0010] Furthermore, in step 3, the conditional temporal convolutional prediction network employs a conditional layer normalization mechanism, using the continuous intent encoding vector as a condition to dynamically generate normalization parameters and modulate the convolutional features.

[0011] Furthermore, in step 3, a two-stage training strategy is adopted during the training phase: the first stage uses only the loss other than the intent-load consistency loss for pre-training; the second stage introduces the intent-load consistency loss for end-to-end fine-tuning. The formula for calculating the intent-load consistency loss is as follows: ; in, For intent-load consistency loss; Encode a continuous intent vector; For future load power sequence; An encoder consisting of multiple stacked fully connected layers is used to encode the continuous intent vector of the input. Projected onto the shared semantic space; This represents a feature extractor composed of temporal convolutional networks, used to extract the input load sequence. Projected onto the shared semantic space; This represents the L2 norm.

[0012] Furthermore, the intent-load consistency loss in step 3 forces the continuous intent encoding vector and the future load prediction sequence to align in the shared semantic space. The calculation method is as follows: after mapping the continuous intent encoding vector and the future load prediction sequence to the same semantic space through the encoder, the Euclidean distance is calculated.

[0013] Furthermore, in step 4, joint training employs a reward function, which includes a basic reward item and an intent fulfillment reward item. The intent fulfillment reward item is a weighted sum of the atomic rewards of each intent and the intent intensity scalar. The intent intensity scalar is obtained by decoding the continuous intent encoding vector through an intent semantic decoder. The intent semantic decoder is a decoder with fixed parameters, used to map the continuous intent encoding vector to the intensity scalar of each intent category.

[0014] Furthermore, the atomic reward for each intention is calculated by the corresponding atomic reward function, which is designed differently according to the functional characteristics of different control agents.

[0015] This invention also provides a vehicle electric balance system that combines spatiotemporal causal graph convolution with intentional load coordination, characterized in that it includes: The multimodal feature extraction module is used to extract feature vectors from multiple modalities. The spatiotemporal causal graph convolutional network module includes a dynamic edge weight generator, stacked spatiotemporal causal graph convolutional layers, and multi-head graph attention pooling layers; The intent-aware spatiotemporal state predictor module includes a causal mask transformer encoder, an intent classification and encoding head, and a conditional temporal convolutional prediction network. The multi-agent collaborative decision-making module includes multiple control agents, an intent semantic decoder, and an experience replay buffer.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes a dynamic edge weight generator and a spatiotemporal causal graph convolutional layer, allowing the fusion weights between modalities to be dynamically adjusted based on the real-time reliability of their respective data. Compared to existing fixed-weighted or simple attention-based fusion methods, this invention exhibits stronger robustness to noisy environments and sensor faults, and achieves higher fusion accuracy.

[0017] This invention introduces an intent-load consistency loss function to force the alignment of continuous intent encoding vectors with future load prediction sequences in a shared semantic space. In existing technologies, intent recognition and load prediction are typically optimized as two independent tasks, lacking semantic connection. This invention constructs a causal reasoning link from "driver intent" to "load change trend," ensuring that load prediction not only meets numerical accuracy requirements but also maintains a high degree of semantic consistency with driver intent, significantly enhancing the forward-looking decision-making capability of the vehicle's electric balance system.

[0018] This invention employs a pseudo-synchronization feature buffering mechanism, replacing complex interpolation or prediction with nearest neighbor search, to solve the problem of multi-source asynchronous data alignment in vehicles with extremely low computational overhead, simplifying the preprocessing process and meeting real-time requirements.

[0019] This invention encodes and decodes continuous intentions into dynamic reward weights, directly linking the optimization objective of multi-agent reinforcement learning to dynamically changing user experiences. Unlike existing technologies that use fixed optimization objectives or discrete mode switching, this invention achieves differentiated and smooth responses from each control agent to complex intentions, driving the system to generate proactive and elegant control strategies. Furthermore, the dominant modality identifier output by multi-head graph attention pooling provides interpretability for the decision-making process. When a sensor fails, the system can automatically adjust the fusion strategy based on changes in attention weights and provide the main control unit with reference clues for fault rollback. Attached Figure Description

[0020] Figure 1 A diagram of the vehicle's electric balance system for coordinating spatiotemporal causal graph convolution and intentional load. Figure 2 A bar chart comparing the accuracy of fixed-weighted fusion, simple attention fusion, and the classification of the present invention's intent; Figure 3 Baseline model and co-training intent - load prediction consistency score box plot; Figure 4 A bar chart comparing the effects of traditional rule-based strategies, fixed-pattern RL strategies, and the vehicle electric balance control of this invention; Figure 5 Convergence comparison curves of independent training and the two-stage training of this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be noted that this embodiment is only a preferred implementation and is not intended to limit the scope of protection of the present invention.

[0022] This invention provides a vehicle electric balance method based on spatiotemporal causal graph convolution and intentional load coordination, comprising the following steps: Step 1: Extract the feature vector with the timestamp closest to the current moment from the feature buffer of each modality at a fixed period, and use it as the pseudo-synchronization input feature of each modality; Step 2: Construct dynamic graph nodes from the pseudo-synchronous input features of each modality, and generate a dynamic soft adjacency matrix through a dynamic edge weight generator; construct a spatiotemporal causal graph convolutional layer, and use the spatiotemporal causal graph convolutional layer to perform spatial aggregation and time propagation of gated recurrent units on the dynamic soft adjacency matrix and the current node features. After stacking multiple layers, obtain global graph features and dominant modality identifiers through multi-head graph attention pooling; cache the global graph features of each historical moment in chronological order as a global graph feature sequence; Step 3: Construct an intent-aware spatiotemporal state predictor. Input the global graph feature sequence into the spatiotemporal state predictor, which encodes the intent state sequence and outputs the intent classification probability and continuous intent encoding vector. Using the intent state sequence and continuous intent encoding vector as conditions, the conditional temporal convolutional prediction network of the spatiotemporal state predictor outputs the future load power sequence. During the training phase, a total loss function including intent-load consistency loss is used for optimization. Step 4: Each control agent collects its own physical state vector; constructs the enhanced observation state of each control agent, which includes the physical state vector, the dominant mode identifier obtained in step 2, the continuous intent encoding vector obtained in step 3, and the future load power sequence, for joint training of multi-agent deep reinforcement learning; after joint training, the vehicle outputs control actions according to the current enhanced observation state to realize the vehicle's electric balance management.

[0023] Reference Figure 1 Specifically, in step 1 of this embodiment, primary features are first extracted from the original data of each modality using a lightweight neural network; secondly, the pseudo-synchronization feature buffering mechanism proposed in this invention is used to approximate time alignment of these asynchronously arriving features, providing stable input for the subsequent fusion module. The raw data for each modality is processed through a lightweight feature extraction network, encoding the high-dimensional raw data into low-dimensional, dense feature vectors. All network outputs are uniformly set to 128-dimensional features to ensure consistent input dimensions for subsequent modules. Visual feature extraction network: A lightweight 7-layer network based on depthwise separable convolution is used as input to a 112×112×3 single-frame RGB image captured by the driver monitoring system, and outputs a 128-dimensional visual feature vector.

[0024] Speech and semantic feature extraction network: It adopts a dual-branch parallel fusion structure. The acoustic branch processes the 1.6-second, 16kHz raw audio waveform, and the semantic branch processes the ASR (Automatic Speech Recognition) text word sequence with a maximum length of 32. The outputs of the two branches are fused to obtain a 128-dimensional speech and semantic joint feature vector.

[0025] Vehicle CAN / Navigation Signal Extraction Network: A stacked gated dilated causal convolutional network is used. The input is vehicle CAN (Controller Area Network) / navigation timing data with a time step of 10 and a feature dimension of 20 per step. The output is a 128-dimensional vehicle dynamic feature vector.

[0026] Environmental feature extraction network: It adopts a multilayer perceptron, takes 5-dimensional environmental scalar data such as temperature and humidity as input, and outputs 128-dimensional environmental feature vector.

[0027] Specifically, in this embodiment, the multimodal raw data is obtained by combining "simulation scene generation with supplementation of non-public datasets".

[0028] Visual images are extracted from driver monitoring videos in non-public driving datasets. The video content covers typical driving states such as fatigue, focus, and distraction. After extracting image frames, a visual feature extraction network is used to extract visual features.

[0029] The voice waveform and text data are generated by an in-vehicle voice command synthesis tool. The command content includes common in-vehicle interaction phrases such as "temperature adjustment", "navigation", and "energy saving mode". The text data is obtained by automatically transcribing the generated voice audio and then extracting semantic features through a voice and semantic feature extraction network.

[0030] The vehicle CAN / navigation signal data utilizes the CARLA simulation engine to construct dynamic driving scenarios, and collects real-time CAN bus and navigation timing data such as vehicle speed, acceleration, steering angle, and GPS path planning.

[0031] Environmental sensor data is simulated in the CARLA simulation environment by setting variable environmental parameters to simulate sensor scalar data under different climatic conditions.

[0032] After all modal data were timestamped, a 10-hour continuous driving time-series dataset was constructed, resulting in more than 300,000 sets of synchronized samples of "multimodal features, driving intention labels, and electrical load power".

[0033] This embodiment employs a pseudo-synchronization feature buffering mechanism, which addresses the data asynchrony problem caused by differences in data acquisition frequency and transmission delay between multiple onboard sensors. This avoids the computational complexity and errors associated with traditional interpolation methods. The present invention proposes a lightweight buffer alignment mechanism: Maintain a first-in-first-out buffer for each mode First-in, first-out buffer Used to store the features output by the feature extraction network for this modality within a recent period and their corresponding timestamps: ; in, Representing each modal feature, v represents visual features, a represents speech and semantic features, n represents vehicle dynamic features, and e represents environmental features; Representing modes In physical time The generated feature vectors, For the corresponding Arrival timestamp; Decision scheduling and alignment operations: The system operates on a fixed decision cycle. In this embodiment, scheduling is performed. This generates a series of decision-making moments. At every decision-making moment For each mode Perform a nearest neighbor search operation from the buffer. Select the one with the closest timestamp The features of this mode serve as pseudo-synchronization input features at the current decision-making moment. This operation can be formally defined as an alignment function. : ; After the above operations, at time Obtain a set of pseudo-synchronous multimodal feature vectors.

[0034] Specifically, in step 2 of this embodiment, the pseudo-synchronous multimodal features output in step 1 are received, and a global feature that integrates multimodal spatiotemporal context is finally output by constructing a dynamic graph and performing spatiotemporal causal graph convolution. Dynamic graph construction: at every decision moment The pseudo-synchronous input features of the four modalities are treated as a fully connected dynamic graph. The four nodes on it.

[0035] Node definition: Each modality corresponds to one node, and the node set is as follows: These represent visual features, speech and semantic features, vehicle dynamic features, and environmental features, respectively. node The initial feature vector is the output of step 1; In this invention, the dynamic adjacency relationship (i.e. the dependency strength between modes) is not a predefined static value, but is generated in real time through a learnable dynamic edge weight generator; In this embodiment, the dynamic edge weight generator is a lightweight multilayer perceptron. Based on the features of all nodes at the current time and the global graph state at the previous time, it calculates a soft adjacency matrix that reflects the dynamic and asymmetric dependencies between modes. : ; ; in, The global graph features represent the previous decision moment, and can be set as the zero vector at the initial moment; Represents a dynamic edge weight generator network; This represents performing Softmax normalization on each row of the matrix to ensure that the sum of the edge weights emanating from any node is 1; This represents the generated dynamic soft adjacency matrix. For matrix elements at time... From the source node mode to target node mode The weight of information transmission strength or credibility; These represent the feature representations of four modalities at the initial moment: visual, speech and semantic, vehicle dynamics, and environment. The spatiotemporal causal graph convolutional layer is the core operation unit of this invention, and they are stacked. layer, For hyperparameters; each layer simultaneously performs spatial (intermodal) aggregation and temporal (historical context) propagation to ensure the spatiotemporal causality of the fusion.

[0036] For the layer( Spatial aggregation (inter-modal fusion): For nodes Aggregate all its neighboring nodes Information. In the fully connected graph setting of this invention, the neighbor set... It includes all other nodes except itself, where, ; in, For nodes In the Characteristics of the layer; Representing the The learnable space transformation weight matrix of the layer, To unify the feature dimension within the spatiotemporal causal graph convolutional network, in this embodiment, it is... ; Dynamic edge weights serve as aggregation weights; The activation function for the Gaussian error linear unit; Representative node Intermediate features after spatial aggregation.

[0037] Temporal propagation (causality guarantee): An independent gated loop unit is maintained for each node to fuse the current spatial aggregation features with the hidden state of the node at the previous time step, ensuring that information processing strictly adheres to temporal causality (does not depend on future information); where, :node At the previous decision moment The hidden state of GRU.

[0038] ; in, This represents a gated loop unit function that takes the current input and the previous hidden state as input and outputs the current hidden state. ; Representing the The learnable time-fusion weight matrix of the layer, The dimension of the GRU hidden layer; The representative layer normalization operation is used to stabilize training; :node In the The final output features of the layer contain multimodal fusion information at the current moment, while also encoding historical temporal context.

[0039] go through After the spatiotemporal causal graph convolutional layer, the features of each node These nodes contain rich multimodal spatiotemporal context information, and it is necessary to aggregate these node features into a global graph-level representation.

[0040] A multi-head graph attention mechanism is used for pooling to dynamically learn the contribution of each modality to the final decision. ; in, The total number of heads that attract attention; For the first Each attention head is assigned to a node. The normalized attention weights reflect the importance of the modality under this attention head; Representing the The learnable projection matrix corresponding to each attention head. Output dimension for each head; This represents a vector concatenation operation; The final output is the global graph feature vector, which serves as the output of the entire multimodal fusion and encoding module; Based on the attention weights, determine which modality's information dominates the fusion decision at the current moment, and add a dominant modality label to that dominant modality: ; in, The dominant modality identifier indicates which information source the fusion result relies on most at the current moment, providing interpretability for system decision-making and a fallback reference in case of specific sensor failure; argmax is the maximum value index operation; Representing the Each attention head is assigned to a node. Normalized attention weights.

[0041] Specifically, in step 3 of this embodiment, the global graph feature sequence output in step 2 is received, and the driver's intention is decoded in parallel by the intention-aware spatiotemporal state predictor to predict the future load. An innovative loss function and training strategy are adopted to ensure that the intention semantics and load pattern are consistent at the high-level semantic level. The input to the intent-aware spatiotemporal state predictor is the nearest The global graph feature sequence at each time step: ; in, For historical context length, it consists of a shared intent-state encoder and two parallel prediction heads; the input sequence is fed into a causal mask transformer encoder, which uses a causal attention mask to ensure that at each time step... The output depends only on Given the previous inputs, strictly adhering to temporal causality, the output is an intention state sequence: ; in, Represents the current moment The intent state vector is condensed up to the current moment, extracting all spatiotemporal context information related to the driver's intent from multimodal information; With the final intention state As input, the intent classification probability is output in parallel through two independent fully connected layers: ; in, The preset number of intent categories; continuous intent encoding vector ,in Dimensions for encoding continuous intent. It is a continuous and dense semantic vector; With the entire intention state sequence and continuous intent encoding vector Given the condition, input the load forecast header to predict the future. The vehicle load power sequence at each time step Its core is a conditional temporal convolutional prediction network; the conditional temporal convolutional prediction network consists of... The system consists of stacked one-dimensional causal dilated convolutional layers, ensuring that predictions rely solely on historical information. This embodiment constructs a conditional injection mechanism to interpret the driver's intent semantics. In the deep learning process for effective load prediction, conditional normalization is employed in the feature processing stage of each convolutional layer, with scaling parameters for the conditional normalization layer. Translation parameters Continuous intent encoding vector Dynamically generated: ; in, This represents a small multilayer perceptron; Scaling parameters for conditional layer normalization; For translation parameters, Encodes a continuous intent vector.

[0042] For the Features output by convolution layers Its conditional normalization operation is as follows: ; in, This indicates element-wise multiplication. These represent the mean and standard deviation of the calculated feature, respectively. To prevent small constants from being divided by zero.

[0043] This invention designs a multi-task joint loss function to train an intent-aware spatiotemporal state predictor and a spatiotemporal causal graph convolutional network. A two-stage training strategy is employed to optimize training stability and final performance. The first stage uses only the loss function excluding the intent-load consistency loss for pre-training. The second stage introduces the intent-load consistency loss and performs end-to-end fine-tuning using continuous sequence data at the vehicle level. The multi-task joint total loss function and the calculation formulas for each loss function are as follows: Intent classification loss: ; in, Representative intention classification loss, supervised intention classification probability With true intent tag Consistency; Intent encoding contrast loss: ; in, Contrast loss for intent encoding, used to bring encodings of similar intents closer together. This extends the encoding of different types of intents to enhance the discriminative power of the encoding space. Respectively represent and A set of sample codes of the same and different classes; The cosine similarity function; Represents temperature parameter; Future load forecasting loss: ; in, Loss prediction for future loads; supervised prediction of load sequences With the actual load sequence Numerical accuracy; Graph structure regularization loss: ; in, Represents the graph structure regularization loss, applied to the dynamic adjacency matrix of spatiotemporal causal graph convolutional networks. This encourages sparse and stable behavior and avoids overfitting. The first term represents the L1 norm, which promotes sparsity; the second term is negative entropy, used to prevent the matrix from becoming excessively sparsity. As weight; Modal decoupling regularization loss: ; in, The modality decoupling regularization loss, after fusion of spatiotemporal causal graph convolutional networks, encourages the high-level features of nodes from different modalities. Maintain a certain degree of independence to improve the model's robustness to single-mode failures; For a shared projection matrix; Intent-load consistency loss, calculation formula: ; in, Intent-load consistency loss is used to enforce continuous intent encoding vectors. With future load power sequence Alignment in high-level semantic space; the encoder For definition in Above, value at The mapping, consisting of multiple stacked fully connected layers, encodes the input continuous intent vector. Projected to a shared semantic space; the feature extractor For definition in Above, value at The mapping, constructed by a temporal convolutional network, maps the input load sequence... Projected onto the shared semantic space; For the dimensions of the shared semantic space; Represents the L2 norm (Euclidean distance); Multi-task loss function: ; in, , , , , Here, represents the weighting coefficients for each loss term, and represents the preset hyperparameters, which are the preset weighting coefficients for each loss term. , , , , 5. This is used to balance different optimization objectives; Furthermore, in this embodiment, step 3 employs a two-stage progressive training method. The first stage uses... The loss combination is used for training; in the first stage, the spatiotemporal causal graph convolutional network learns basic multimodal dynamic fusion, and the intention-aware spatiotemporal state predictor learns basic intention classification and load temporal prediction, but the two tasks are relatively independent. The second phase introduces the core Losses are addressed through end-to-end fine-tuning using continuous long-term time-series data at the vehicle level; this stage leverages existing capabilities to... Strongly coupling the intent header and the load prediction header drives the network to learn the deep semantic relationship between intent and load changes, achieving true collaborative prediction; This embodiment uses a workstation equipped with an Intel Core i9-13900K processor, an NVIDIA RTX 4090 graphics card, and 64GB of RAM. The algorithm is implemented based on Python 3.9 and PyTorch 2.0 frameworks. The vehicle electrical balance simulation relies on the MATLAB / Simulink and AMESim joint simulation platform. The 10-hour dataset is divided into training, validation, and test sets in an 8:1:1 ratio, with a total of 100 training rounds. The performance of each method is tested under "clear scenes" and "noisy scenes" (simulating sensor noise or partial modal loss), as well as under different driving intentions (comfort mode, economy mode, and sport mode).

[0044] Specifically, in step 4 of this implementation, the intention and prediction information output in step 3 are used to construct a multi-agent deep reinforcement learning system. Through the intention-driven reward shaping mechanism, the collaborative optimization of each control agent is realized, and finally the vehicle electric balance management that meets the driver's dynamic intention is achieved. For each control agent Construct an enhanced observation state Injecting holistic understanding and future information to facilitate decision-making: ; in, For intelligent agents Its own traditional physical state vector; The continuous intent encoding vector from step 3 provides unified, high-level, and continuous semantic target guidance for all agents; Represents the future load power sequence predicted from step 3. In the middle, with intelligent agents A future segment strongly correlated with management workload; Length is ,generally ; The dominant mode identifier comes from step 2; This invention directly links the optimization objective of reinforcement learning with the dynamically changing "user experience" through reward shaping, and transforms the abstract intention semantics into quantifiable reward weights through an intention semantic decoder with fixed parameters. Through a pre-defined, parameter-fixed intent semantic decoder The intent semantic decoder is typically a shallow fully connected network that encodes continuous intent vectors. Decoded into a set of interpretable intent strength scalars : ; in, This represents a set of intent categories, in this embodiment. ; Representation Intent The intensity scalar at the current moment satisfies Quantify the relative importance of each category in the driver's complex intentions; A hybrid reward function driven by design intent, for each agent At any moment Instant rewards The design consists of three parts: ; Among them, basic reward items Rewards are given to ensure the basic safety and stability of the system, such as penalties for voltage exceeding limits and current exceeding limits; one of the innovations of this invention lies in its intention to satisfy these reward items. For intelligent agents Intent A specially designed atomic reward function transforms abstract intentions into specific physical metrics related to the agent's actions and states; The dynamic weights determine the atomic reward at the current moment. The proportion of total reward; the higher the intent intensity, the greater the weight of the corresponding optimization objective in the decision-making process; collaborative reward items. Encourage collaborative behavior among intelligent agents and reward improvements in overall energy efficiency.

[0045] In this embodiment, the atomic reward function includes comfort intention, economic intention, and movement intention; Comfort intention ( Applications in air conditioner controllers: ; in, , All are weighted coefficients, with the first item being the cabin temperature. Approaching the set temperature The second penalty is the compressor power. drastic changes , To reduce noise and vibration; Economic intentions ( Applied to battery manager: ; in, , All are weighting coefficients. and These represent the amplitude and rate of change of the penalty discharge current, respectively. Encourage gentle discharge to extend battery life; Motion intention ( Applications to vehicle energy management systems: ; in, , All are weighting coefficients; This refers to the bus voltage. This is the rated bus voltage; A reward will be given when the system prioritizes power supply to the power system, while allowing bus voltage... Within a safe range, there is a greater dynamic range, and the penalty term weighting... It can be reduced.

[0046] The training process for multi-agent collaborative decision-making adopts a centralized training-decentralized execution framework. In this embodiment, the MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm is used.

[0047] Each agent interacts with the environment according to its current strategy, and updates its current augmented observation state. Actions to be performed Instant rewards received and the enhanced observation status at the next moment. As a lesson learned, it is stored in a shared experience replay buffer; during training, the central commentator network can access the augmented observation states of all agents. and actions To learn and evaluate the overall value of joint actions or state value ; Policy updates are performed using gradient ascent to update the policy network of each agent. To maximize the expected long-term cumulative discount rewards After online training is completed, deploy the policy network for each agent. ; At every decision moment The system processes the data according to steps 1-3 to obtain... For each intelligent agent Construct its enhanced observation state Each intelligent agent Policy network according to Independent output control action Actions of all intelligent agents Together they act on the vehicle's electrical system to achieve coordinated and optimized vehicle electrical balance management that meets the driver's intentions; ultimately, they output a set of coordinated control actions that directly drive each electrical load controller.

[0048] Specifically, this embodiment designs multiple baseline methods for comparison, covering four dimensions: multimodal fusion accuracy, intent-load consistency, power balance control effect, and training convergence. The comparison methods include: Baseline Method 1 (Fixed Weighted Fusion): Multimodal features are assigned fixed weights and then concatenated before being input into the classifier.

[0049] Baseline Method 2 (Simple Attention Fusion): Employs a traditional scaled dot product attention mechanism for weighted fusion of multimodal features.

[0050] Baseline Method 3 (Independently Trained Model): The intent classification network and the load prediction network are trained separately, without applying intent-load consistency loss (excluding...). loss).

[0051] Traditional rule-based strategy: power balance controller based on preset thresholds.

[0052] In contrast to RL methods (fixed-mode RL strategies), the intent is input as a discrete mode switch, rather than the continuous intent encoding vector proposed in this invention.

[0053] Each method was trained and evaluated on the same training, validation and test sets, and the intent classification accuracy, intent-load consistency score, voltage fluctuation RMS (Root Mean Square Voltage), intent satisfaction, overall satisfaction and training loss curve were recorded.

[0054] Specifically, regarding the accuracy of intent classification, refer to Figure 2In clear scenes, the classification accuracy of the proposed method reaches 94.8%, significantly higher than that of fixed-weighted fusion (82.7%) and simple attention fusion (90.2%). In noisy scenes, the accuracy of the proposed method remains at 93.1%, while that of fixed-weighted fusion drops to 78.3% and simple attention fusion drops to 92.5%. The results indicate that the dynamic fusion mechanism proposed in this invention has stronger robustness to sensor noise and modality loss.

[0055] Reference Figure 3 The method of this invention introduces an intent-load consistency loss through collaborative training. The median semantic consistency score between the continuous intent encoding vector and the predicted load sequence is approximately 0.88, and the distribution is concentrated. In contrast, the median score of the independently trained model is only 0.65, and the fluctuation is large. This proves that the loss effectively solves the "semantic disconnect" problem between intent understanding and load control.

[0056] Reference Figure 4 Regarding voltage fluctuation (RMS), this invention achieves 0.12, superior to the traditional rule-based strategy (0.18) and the fixed-mode RL strategy (0.15). In terms of intent satisfaction, this invention reaches 0.95 in comfort mode and 0.92 in economy mode, both significantly higher than the comparative methods. The overall satisfaction score is 0.88, significantly leading the comparison. This demonstrates that the intent-driven multi-agent collaborative control strategy can proactively and smoothly satisfy dynamic driving intentions.

[0057] Reference Figure 5 In the two-stage training strategy employed in this invention, the loss has been reduced to a low level after the first 40 rounds of pre-training, and the second stage introduces... After the initial loss, the total loss further decreased to approximately 0.2, indicating stable convergence. In contrast, the independently trained model converged slowly, with a final loss of approximately 0.4. This demonstrates that two-stage training balances the stability of pre-training with end-to-end synergy, resulting in higher training efficiency.

[0058] This embodiment also provides a vehicle electric balance system that coordinates spatiotemporal causal graph convolution with intentional load, including: The multimodal feature extraction module is used to extract feature vectors from multiple modalities. The spatiotemporal causal graph convolutional network module includes a dynamic edge weight generator, stacked spatiotemporal causal graph convolutional layers, and multi-head graph attention pooling layers; The intent-aware spatiotemporal state predictor module includes a causal mask transformer encoder, an intent classification and encoding head, and a conditional temporal convolutional prediction network. The multi-agent collaborative decision-making module includes multiple control agents, an intent semantic decoder, and an experience replay buffer.

[0059] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for vehicle electrical balancing based on spatiotemporal causal graph convolution and intentional load coordination, characterized in that, Includes the following steps: Step 1: Extract the feature vector with the timestamp closest to the current moment from the feature buffer of each modality at a fixed period, and use it as the pseudo-synchronization input feature of each modality; Step 2: Construct dynamic graph nodes from the pseudo-synchronous input features of each modality, and generate a dynamic soft adjacency matrix using a dynamic edge weight generator; construct a spatiotemporal causal graph convolutional layer, and use the spatiotemporal causal graph convolutional layer to perform spatial aggregation and time propagation of gated recurrent units on the dynamic soft adjacency matrix and the current node features. After stacking multiple layers, obtain global graph features and dominant modality identifiers through multi-head graph attention pooling; construct a global graph feature sequence composed of global graph features from each historical time step. Step 3: Construct an intent-aware spatiotemporal state predictor; input the global graph feature sequence into the spatiotemporal state predictor, encode it to obtain an intent state sequence, and output the intent classification probability and continuous intent encoding vector; using the intent state sequence and continuous intent encoding vector as conditions, output the future load power sequence through the conditional temporal convolutional prediction network of the spatiotemporal state predictor; wherein, during the training phase, a total loss function including intent-load consistency loss is used for optimization; Step 4: Each control agent collects its own physical state vector; constructs the enhanced observation state of each control agent, which includes the physical state vector, the dominant mode identifier, the continuous intent encoding vector, and the future load power sequence, for joint training of multi-agent deep reinforcement learning; after joint training, the vehicle outputs control actions according to the current enhanced observation state to realize the vehicle's electric balance management.

2. The vehicle electrical balancing method according to claim 1, characterized in that, The dynamic edge weight generator mentioned in step 2 is a multilayer perceptron. The input to the dynamic edge weight generator is the concatenation of the features of all current nodes and the global graph features from the previous time step. ； ； in, The global graph features representing the previous decision-making moment; Represents a dynamic edge weight generator network; This represents performing Softmax normalization on each row of the matrix; This represents the generated dynamic soft adjacency matrix. For matrix elements at time... From the source node mode to target node mode The weight of information transmission strength or credibility; These represent the feature representations of four modalities at the initial moment: visual, speech and semantic, vehicle dynamics, and environment. To output a normalized dynamic soft adjacency matrix.

3. The vehicle electrical balancing method according to claim 1, characterized in that, The spatiotemporal causal graph convolutional layer described in step 2 includes a graph convolution module and a gated recurrent unit. Spatial aggregation is based on the weighted fusion of neighbor node features using a dynamic soft adjacency matrix and mapped by an activation function. Temporal propagation uses the gated recurrent unit to fuse the current aggregated features with the hidden state of the previous time step.

4. The vehicle electrical balancing method according to claim 1, characterized in that, In step 2, the multi-head graph attention pooling, while outputting global graph features, determines the dominant mode identifier at the current moment by summing the attention weights of each mode and taking the mode corresponding to the maximum value.

5. The vehicle electrical balancing method according to claim 1, characterized in that, In step 3, the conditional temporal convolutional prediction network adopts a conditional layer normalization mechanism, which uses the continuous intent encoding vector as a condition to dynamically generate normalization parameters and modulate the convolutional features.

6. The vehicle electrical balancing method according to claim 1, characterized in that, Step 3 employs a two-stage training strategy: the first stage uses only the loss excluding the intent-load consistency loss for pre-training; the second stage introduces the intent-load consistency loss for end-to-end fine-tuning. The formula for calculating the intent-load consistency loss is as follows: ； in, For intent-load consistency loss; Encode a continuous intent vector; For future load power sequence; An encoder consisting of multiple stacked fully connected layers is used to encode the continuous intent vector of the input. Projected onto the shared semantic space; This represents a feature extractor composed of temporal convolutional networks, used to extract the input load sequence. Projected onto the shared semantic space; This represents the L2 norm.

7. The vehicle electrical balancing method according to claim 1, characterized in that, The intent-load consistency loss described in step 3 forces the continuous intent encoding vector and the future load prediction sequence to align in the shared semantic space. The calculation method is as follows: after mapping the continuous intent encoding vector and the future load prediction sequence to the same semantic space through the encoder, the Euclidean distance is calculated.

8. The vehicle electrical balancing method according to claim 1, characterized in that, In step 4, joint training employs a reward function, which includes a basic reward item and an intent fulfillment reward item. The intent fulfillment reward item is a weighted sum of the atomic rewards of each intent and the intent intensity scalar, and the intent intensity scalar is obtained by decoding the continuous intent encoding vector through the intent semantic decoder. The intent semantic decoder is a decoder with fixed parameters, used to map continuous intent encoding vectors into intensity scalars for each intent category.

9. The vehicle electrical balancing method according to claim 8, characterized in that, The atomic reward for each intention is calculated by the corresponding atomic reward function, which is designed differently according to the functional characteristics of different control agents.

10. A vehicle electric balancing system that combines spatiotemporal causal graph convolution with intentional load coordination, applied to the vehicle electric balancing method according to any one of claims 1 to 9, characterized in that, include: The multimodal feature extraction module is used to extract feature vectors from multiple modalities. The spatiotemporal causal graph convolutional network module includes a dynamic edge weight generator, stacked spatiotemporal causal graph convolutional layers, and multi-head graph attention pooling layers; The intent-aware spatiotemporal state predictor module includes a causal mask transformer encoder, an intent classification and encoding head, and a conditional temporal convolutional prediction network. The multi-agent collaborative decision-making module includes multiple control agents, an intent semantic decoder, and an experience replay buffer.

Images