Fuel cell vehicle range prediction system based on historical real-time energy consumption

Abstract

The application discloses a kind of fuel cell vehicle driving range prediction systems based on historical real-time energy consumption, wherein multi-source heterogeneous data perception layer collects data to construct multi-modal data pool, and uses hidden Markov model output driving style determination result;Dynamic response optimization layer is based on model predictive control algorithm dynamically adjusts the output power of electric pile and BOP component;Adaptive prediction model layer aggregates vehicle end and cloud parameters using federated learning framework, combines gated recurrent unit network to dynamically predict energy consumption, and uses convolutional neural network and range rolling prediction model to generate range prediction value and energy consumption distribution heat map;User interaction feedback layer displays prediction results and intelligent recommended path through three-dimensional visualization interface, and returns user feedback signal to front end each layer to realize system optimization.The system of the present application significantly improves the accuracy and system adaptive ability of fuel cell vehicle driving range prediction through multi-layer cooperative closed-loop control mechanism.

Description

Technical Field

[0001] This invention belongs to the field of driving range prediction technology, specifically relating to a fuel cell vehicle driving range prediction system based on historical real-time energy consumption. Background Technology

[0002] Driven by the global goal of carbon neutrality, fuel cell vehicles have become a core pathway for achieving low-carbon transformation in the medium and heavy-duty commercial vehicle sector. Leveraging my country's comprehensive industrial chain, fuel cell systems have reached international leading levels in core indicators such as membrane electrode assembly (MEA) technology and stack power density, laying a solid industrial foundation for the large-scale promotion of this technology in the medium and heavy-duty commercial vehicle sector. Compared to pure electric vehicles, fuel cell vehicles possess significant advantages: shorter refueling times, longer driving ranges on a single charge, and emissions consisting solely of water vapor throughout operation, perfectly meeting the high-frequency, long-distance operational needs of commercial vehicles. Based on these technological characteristics and market demands, the research and application of fuel cell vehicle range prediction systems are particularly important.

[0003] However, the current fuel cell vehicle range prediction system suffers from a significant problem of dynamic response lag. The output power of a fuel cell is affected by multiple factors such as temperature, humidity, and air pressure. Existing range prediction systems struggle to capture power fluctuations in real time. For example, during rapid acceleration, battery response delays may cause the predicted value to deviate from the actual energy consumption by more than 15%. Therefore, a fuel cell vehicle range prediction system based on historical real-time energy consumption is proposed. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a fuel cell vehicle range prediction system based on historical real-time energy consumption. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] A fuel cell vehicle range prediction system based on historical real-time energy consumption includes a multi-source heterogeneous data sensing layer for forming closed-loop control, a dynamic response optimization layer, an adaptive prediction model layer, and a user interaction feedback layer.

[0006] The multi-source heterogeneous data perception layer is configured to collect vehicle physical parameters of the fuel cell, dynamic information of the environment around the vehicle, and driving style data in the current control cycle, and combine the three into a multimodal data pool; use the deployed vehicle edge computing nodes to preprocess the raw data in the multimodal data pool to obtain multimodal feature data, and output the driving style determination result based on the loaded hidden Markov model.

[0007] The dynamic response optimization layer is configured to dynamically adjust the output power of the fuel cell stack and BOP component based on the multimodal feature data and driving style determination results in the current control cycle through a model predictive control algorithm, and combine the vehicle edge computing node to make localized real-time decisions to generate control adjustment commands and corresponding system status data.

[0008] The adaptive prediction model layer is configured to aggregate model parameters from the vehicle and cloud through a federated learning framework during the current control cycle, and dynamically predict energy consumption data containing the control adjustment commands and system state data by combining a gated recurrent unit (GRU) network, so as to adjust the control adjustment commands; based on the dynamic prediction results of energy consumption data and environmental dynamic information, and by using a convolutional neural network (CNN) and a rolling prediction model for driving range, a rolling prediction value for driving range and a corresponding energy consumption distribution heatmap for a future period of time are generated.

[0009] The user interaction feedback layer is configured to display the rolling prediction value of the driving range and the energy consumption distribution heat map through a three-dimensional visualization interface during the current control cycle. It also displays the planned priority path based on real-time traffic data and vehicle status dynamic path planning using an intelligent path recommendation algorithm. The feedback signal generated by the user based on the displayed information is sent back to the multi-source heterogeneous data perception layer to correct the driving style data, and then sent back to the dynamic response optimization layer to readjust the output power control.

[0010] Beneficial effects:

[0011] This application utilizes onboard edge computing nodes to perform localized preprocessing and feature extraction of raw data from a multimodal data pool. It then employs model predictive control algorithms to dynamically adjust the output power of the fuel cell stack and BOP components, achieving millisecond-level localized real-time decision-making and significantly improving system response speed. Simultaneously, by adopting a federated learning framework and a lightweight online model update mechanism, it significantly reduces cloud communication latency while ensuring data privacy, guaranteeing the real-time performance and accuracy of the prediction model. Finally, through closed-loop feedback signal transmission, it continuously corrects driving style data and power control strategies, enabling the system to possess continuous self-optimization capabilities. This effectively solves the dynamic response lag problem of existing fuel cell vehicle range prediction systems, significantly improving prediction accuracy and system reliability.

[0012] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0013] Figure 1 This is a structural diagram of a fuel cell vehicle range prediction system based on historical real-time energy consumption proposed in this invention.

[0014] Figure 2This is a system block diagram of a fuel cell vehicle range prediction system based on historical real-time energy consumption proposed in this invention.

[0015] Figure 3 This is a diagram of the hierarchical data processing architecture of a fuel cell vehicle range prediction system based on historical real-time energy consumption proposed in this invention.

[0016] Figure 4 This is a flowchart of the federated learning collaborative mechanism of a fuel cell vehicle range prediction system based on historical real-time energy consumption proposed in this invention.

[0017] Figure 5 This is a feedback control closed-loop flowchart of a fuel cell vehicle range prediction system based on historical real-time energy consumption proposed in this invention.

[0018] Figure 6 This is a flowchart of a multimodal decision-making routing system for predicting the driving range of a fuel cell vehicle based on historical real-time energy consumption, as proposed in this invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0020] Combination Figures 1 to 6 This application provides a fuel cell vehicle range prediction system based on historical real-time energy consumption, comprising: a multi-source heterogeneous data sensing layer forming a closed-loop control, a dynamic response optimization layer, an adaptive prediction model layer, and a user interaction feedback layer;

[0021] The multi-source heterogeneous data perception layer is configured to collect vehicle physical parameters of the fuel cell, dynamic information of the environment around the vehicle, and driving style data in the current control cycle, and combine the three into a multimodal data pool; use the deployed vehicle edge computing nodes to preprocess the raw data in the multimodal data pool to obtain multimodal feature data, and output the driving style determination result based on the loaded hidden Markov model.

[0022] This application constructs a multimodal data pool by real-time acquisition of vehicle physical parameters such as fuel cell voltage, current, hydrogen pressure, temperature, and battery pack temperature, and integrates environmental dynamic information such as road slope, rainfall, and road surface friction coefficient obtained from cameras, radar, and V2X communication. It also combines steering wheel torque, accelerator pedal opening, and user driving style captured by a high-precision positioning module. Simultaneously, it deploys onboard edge computing nodes to perform localized preprocessing and feature extraction on the raw data in the multimodal data pool, reducing cloud communication latency, enabling rapid data response, and sending multimodal data to the dynamic response optimization layer, thereby providing a low-latency, high-precision input source for the upper dynamic response optimization layer.

[0023] The dynamic response optimization layer is configured to dynamically adjust the output power of the fuel cell stack and BOP component based on the multimodal feature data and driving style determination results in the current control cycle through a model predictive control algorithm, and combine the vehicle edge computing node to make localized real-time decisions to generate control adjustment commands and corresponding system status data.

[0024] This application dynamically adjusts the output power of the fuel cell stack and BOP component through the model predictive control (MPC) algorithm, and performs localized real-time decision-making in conjunction with the on-board edge computing node. It sends optimized fuel cell power commands (such as fuel cell stack output power and BOP component control signals) and status (such as hydrogen consumption rate and battery pack temperature) to the adaptive predictive model layer, thereby significantly shortening the power response time to within 50ms.

[0025] The adaptive prediction model layer is configured to aggregate model parameters from the vehicle and cloud through a federated learning framework during the current control cycle, and dynamically predict energy consumption data containing the control adjustment commands and system state data by combining a gated recurrent unit (GRU) network, so as to adjust the control adjustment commands; based on the dynamic prediction results of energy consumption data and environmental dynamic information, and by using a convolutional neural network (CNN) and a rolling prediction model for driving range, a rolling prediction value for driving range and a corresponding energy consumption distribution heatmap for a future period of time are generated.

[0026] This application aggregates model parameters between the vehicle and the cloud using a federated learning framework, combines a GRU network to dynamically model time-series energy consumption data (such as historical hydrogen consumption and power output sequences), and utilizes CNN to extract environmental image features (such as road congestion heatmaps and rainfall intensity distribution) to generate rolling predictions of driving range and energy consumption distribution heatmaps for a future period, which are then output to the user interaction feedback layer. Through a lightweight online model update mechanism, cloud communication latency is reduced, shortening the prediction response time. At the same time, based on real-time traffic data and vehicle status, intelligent route planning suggestions (such as avoiding congested road sections and recommending hydrogen refueling stations) are dynamically generated to support users in quickly adjusting their driving strategies to adapt to sudden acceleration and extreme weather conditions.

[0027] The user interaction feedback layer is configured to display the rolling prediction value of the driving range and the energy consumption distribution heat map through a three-dimensional visualization interface during the current control cycle. It also displays the planned priority path based on real-time traffic data and vehicle status dynamic path planning using an intelligent path recommendation algorithm. The feedback signal generated by the user based on the displayed information is sent back to the multi-source heterogeneous data perception layer to correct the driving style data, and then sent back to the dynamic response optimization layer to readjust the output power control.

[0028] This application uses a 3D visualization interface to display the remaining hydrogen, predicted driving range, and energy consumption distribution heatmap in real time. Combined with an intelligent route recommendation algorithm, it dynamically plans the optimal route based on real-time traffic data and vehicle status, enabling users to intuitively perceive changes in driving range and quickly adjust their driving strategies. This alleviates range anxiety caused by information lag. At the same time, feedback signals are sent back to the multi-source heterogeneous data perception layer (to correct driving style data) and the dynamic response optimization layer (to adjust power control strategies).

[0029] In one specific embodiment of this application, the multi-source heterogeneous data sensing layer is specifically configured as follows:

[0030] S110, during the current control cycle, collects fuel cell voltage, current, hydrogen pressure, temperature, and battery pack temperature in real time as vehicle physical data;

[0031] S120, during the current control cycle, acquires road slope, rainfall, road surface friction coefficient, images of the road ahead of the vehicle, traffic signs, and dynamic information of pedestrians and vehicles as environmental dynamic data through cameras, radar and V2X communication.

[0032] This step can also obtain road congestion index (TCI), accident hotspot locations, and construction section information through the V2X communication module, with a data update frequency of 1Hz; it can also obtain key parameters such as remaining hydrogen (SOC_H2), predicted driving range (Range_pred), and battery pack temperature (T_batt) from the multi-source heterogeneous data perception layer, combined with the lane curvature identified by the camera and the speed of surrounding vehicles detected by the radar.

[0033] S130, in the current control cycle, driving style data captured by the steering wheel torque, accelerator pedal opening, and high-precision positioning module; wherein, the driving style data includes steering wheel torque signal, accelerator pedal opening data, and vehicle trajectory data;

[0034] This step can collect steering operation force and frequency data through the steering wheel torque sensor, and obtain the accelerator pedal depth change curve by combining it with the accelerator pedal opening sensor; use a high-precision positioning module (such as GPS+IMU) to record vehicle trajectory characteristics, including the frequency of rapid acceleration and deceleration, and the turning radius change pattern; and simultaneously collect the longitudinal acceleration and yaw rate dynamic parameters from the vehicle CAN bus data.

[0035] S140, the vehicle physical data, the dynamic environment data, and the driving style data are combined into a multimodal data pool;

[0036] S150, the vehicle-mounted edge computing node is used to clean, standardize, and extract features from the raw data in the multimodal data pool to obtain multimodal feature data;

[0037] Edge computing nodes receive raw data from the multi-source heterogeneous data perception layer, including vehicle physical parameters, environmental dynamic information, and driving style data; they clean the raw data, remove noise and outliers, and perform standardization to improve data quality; they extract key features, such as driving style features (extracted through data such as steering wheel torque and accelerator pedal opening), environmental parameter features (such as road slope and rainfall), and vehicle status features (such as fuel cell voltage and current).

[0038] This application deploys vehicle-mounted edge computing nodes to perform localized preprocessing and feature extraction on raw data in a multimodal data pool, reducing cloud communication latency, enabling rapid data response, and sending multimodal data to the dynamic response optimization layer, thereby providing a low-latency, high-precision input source for the upper dynamic response optimization layer.

[0039] S160, the hidden Markov model is loaded using the vehicle edge computing node, and the feature data corresponding to the driving style data is used as input to obtain the output driving style determination result.

[0040] In this application, the vehicle-mounted edge computing node loads a pre-trained machine learning model, such as a Hidden Markov Model (HMM) for driving style recognition, uses the extracted features as model input, performs real-time inference, and generates driving style labels.

[0041] In one specific embodiment of this application, S160 includes:

[0042] S161, Perform FFT transformation on the steering wheel torque feature in the multimodal feature data to extract the energy proportion of the predetermined frequency band;

[0043] This application performs sliding window framing processing on the steering wheel torque signal, with a window length of 5 seconds and an overlap rate of 50%, to preserve the temporal characteristics of driving behavior; it uses a Kalman filter algorithm to denoise the steering wheel torque signal to eliminate measurement errors caused by road bumps; and it normalizes the accelerator pedal opening data, mapping the 0-100% opening to the 0-1 normalized range.

[0044] S162, calculates the standard deviation of the rate of change of the accelerator pedal opening characteristic per unit time;

[0045] S163, Based on the frequency band energy ratio and the standard deviation of the rate of change, a driving style feature vector containing the probabilities of aggressive, normal, and conservative driving styles is generated through a hidden Markov model.

[0046] S164, Select the driving style with the highest probability from the driving style feature vector as the predicted driving style, and use it and the corresponding probability as the driving style determination result.

[0047] This application calculates the standard deviation of the accelerator pedal opening change rate and the peak value of the steering wheel angular velocity per unit time; performs FFT transformation on the steering wheel torque signal to extract the energy proportion of the 0.5-2Hz frequency band (characterizing the steering operation frequency), and statistically analyzes the frequency and duration distribution of rapid acceleration events (pedal opening change rate > 2% / s); adopts an improved Hidden Markov Model (HMM) for behavior pattern recognition, and sets three hidden states: aggressive, normal, and conservative; the training dataset contains 1000 hours of real-vehicle data from multiple drivers, and typical driving segments are labeled through semi-supervised learning; the model output is a three-dimensional probability vector, i.e., a driving style feature vector, and the maximum probability value is taken as the style determination result.

[0048] This application also sets a sliding time window (T=30s) for online learning, and triggers model updates when the change in the probability distribution of driving style exceeds the threshold (ΔP>0.3); adopts an incremental learning strategy, only updates the relevant state transition matrix parameters, and avoids global model retraining; encapsulates the recognition results in JSON format, including style labels, confidence scores and feature vectors, and sends them to the dynamic response optimization layer via CAN FD bus.

[0049] In one specific embodiment of this application, the dynamic response optimization layer is specifically configured as follows:

[0050] S210, Establish a dynamic physical model of the fuel cell stack and BOP assembly;

[0051] This step establishes a dynamic physical model of the fuel cell stack and BOP components (such as air compressor, humidifier, and hydrogen circulation pump), including electrochemical reaction kinetics, thermodynamic equilibrium, and fluid dynamics equations.

[0052] S220, using the unscented Kalman filter algorithm, the key state variables in the dynamic physical model are estimated in real time to obtain the estimated values ​​of the key state variables;

[0053] This step uses the Unscented Kalman Filter (UKF) algorithm to estimate key state variables in real time, including parameters that are difficult to measure directly, such as the internal humidity of the fuel cell stack, the water content of the membrane electrode, and the temperature distribution of the bipolar plates.

[0054] S230, Discretize the dynamic physical model to obtain the discretized state space equation, and construct a rolling prediction model based on the discretized state equation;

[0055] This step is based on discretized state-space equations to construct a rolling prediction model with a prediction time domain of Np=100ms and a control time domain of Nc=50ms. The inputs include measurable variables such as power demand, hydrogen pressure, air flow rate, and coolant temperature, and the outputs include stack output voltage, BOP component power consumption, and key efficiency performance indicators.

[0056] S240 determines the required power based on the driving style assessment result, and constructs a constrained objective function based on the required power, expressed as:

[0057] 1( P req( k ) P stack( k ) w 2Δ u ( k w 3 T stack( k

[0058] Among them, the weighting coefficient 1 = 0.7 2 = 0.2, 3 = 0.1 req is the power demand. T stack represents the temperature distribution of the bipolar plates, Δ u ( k) The rate of change of the control quantity at time k is given by the following constraints: hydrogen utilization rate less than 0.8 (to prevent membrane electrode starvation), air stoichiometry greater than 1.5 (to ensure oxygen supply), and coolant temperature gradient less than 5℃ / s (to avoid thermal shock).

[0059] S250, in the current control cycle, the rolling prediction model is loaded using the vehicle-mounted edge computing node, taking the required power and the multimodal feature data as inputs, and the objective function as the objective of the rolling prediction model, so that it outputs the predicted value of the stack output voltage and the predicted value of the BOP component power consumption in the future time domain.

[0060] In each control cycle (Δt=10ms), the constrained objective function is solved using the interior-point method. Solving the objective function yields the optimal sequence of the stack output power and BOP component control signals in the future time domain. The aim is to achieve real-time optimized control of the fuel cell vehicle range prediction system by dynamically adjusting the power output while meeting constraints such as power point tracking accuracy and thermal management safety.

[0061] S260, Based on the predicted output voltage of the fuel cell stack and the predicted power consumption of the BOP component, control adjustment commands and driving strategy suggestions are generated;

[0062] The onboard edge computing node loads a Model Predictive Control (MPC) model for power control, i.e., a rolling prediction model, to generate predicted values ​​for the fuel cell stack output voltage and the power consumption of the BOP (Balance of Plant) component. Combined with current vehicle status and environmental information, the edge computing node generates real-time decisions. During decision generation, safety constraints and performance optimization objectives are considered to ensure hydrogen utilization remains within a safe range and optimize system efficiency. These decisions include power adjustment commands (such as adjusting the output power of the fuel cell stack and BOP component) or driving strategy suggestions (such as suggested accelerator pedal opening and steering angle).

[0063] S270, the control adjustment command, driving strategy suggestion and corresponding vehicle system status data are sent to the corresponding BOP component actuator via the CAN bus.

[0064] The onboard edge computing node sends the first element of the control and adjustment command sequence (air compressor speed command, hydrogen circulation pump flow command, fuel cell stack operating temperature setpoint) to the actuator via the CAN bus. In the next control cycle, the latest measured values ​​are used to perform feedback correction on the rolling prediction model, correcting the parameter deviations of the rolling prediction model, specifically including: fuel cell stack internal resistance correction. R stack= kp ( V meas V pred)+ ki ∫( V meas V pred) ;in, meas is the voltage measurement value. pred is the predicted voltage value; mass transfer coefficient correction amount: km = T stack P H2) (Dynamically adjusted based on table lookup method).

[0065] When a load mutation (Δ) is detected When req / Δt>5kW / s, the pre-adjustment mode is activated, increasing the air compressor speed reserve 30ms in advance, pre-opening the hydrogen bypass valve, and increasing the coolant flow rate; when hydrogen leakage occurs, the emergency shutdown procedure is immediately executed, the main hydrogen valve is closed, nitrogen purging is started, and the audible and visual alarms and remote diagnostic interface are triggered.

[0066] In one specific embodiment of this application, the adaptive prediction model layer is specifically configured as follows:

[0067] S310: Using local data from each vehicle, a local model is trained on the vehicle to obtain local model update parameters. The local model update parameters are then encrypted and transmitted to the cloud via a secure communication protocol. All local model update parameters are aggregated on the cloud to train a global model and obtain a new global model. The new global model update is then distributed to the vehicle via a secure communication protocol. The new global model is then retrained on the vehicle using local data.

[0068] In this application, the vehicle-side uses local data (such as driving behavior data, environmental data, vehicle state data, etc.) for model training. During the training process, the model parameters are updated based on the local data to generate local model updates (such as parameter gradients). Local training uses lightweight algorithms, such as stochastic gradient descent (SGD), to adapt to the limitations of onboard computing resources.

[0069] Combination Figure 3 and Figure 4 The vehicle-side encrypts local model updates to ensure privacy and security during data transmission. The encrypted model updates are then uploaded to the cloud-based federated learning platform via secure communication protocols (such as TLS / SSL). The cloud-based federated learning platform collects encrypted model updates from multiple vehicle-side endpoints and merges these updates using an aggregation algorithm (such as federated averaging) to generate a global model update. This aggregation process considers the differences in computing power and data distribution among the vehicle-side endpoints to improve the fairness and generalization ability of the global model. The cloud then decrypts the aggregated global model update and generates a new global model. This new global model is distributed back to the vehicle-side endpoints via a secure communication protocol, serving as the starting point for the next round of local training. The vehicle-side endpoints use the received global model for the next round of local training, repeating the above steps. As the number of iterations increases, the global model gradually integrates data features from multiple vehicle-side endpoints, improving prediction accuracy and generalization ability.

[0070] S320: The gated recurrent unit (GRU) network is used to dynamically predict the time-series energy consumption data containing the control and regulation commands and system status data, so as to adjust the control and regulation commands.

[0071] The S330, based on dynamic modeling results and environmental dynamic information, generates a rolling prediction value of the driving range and the corresponding energy consumption distribution heat map for a future period of time by using a convolutional neural network (CNN) and a rolling prediction model of the driving range.

[0072] In one specific embodiment of this application, S320 includes:

[0073] S321 collects historical energy consumption data of vehicles, preprocesses it, and then divides it into training set, validation set, and test set;

[0074] This application collects historical energy consumption data, including time-series signals such as power demand, speed, acceleration, and ambient temperature during vehicle operation. The data is cleaned, missing and outlier values ​​are removed to ensure data quality, and feature engineering is performed to extract time-series features, such as average power, maximum acceleration, and rate of change of speed within a sliding window. The preprocessed data is then divided into training, validation, and test sets.

[0075] S322, the pre-built GRU network is trained using the training set and validated using the trained GRU network using the validation set, so as to readjust the network parameters to obtain the trained GRU network.

[0076] refer to Figure 5 Define a GRU network structure consisting of an input layer, GRU layers, and an output layer. Set the parameters of the GRU layers, such as the number of hidden units and the activation function (e.g., tanh or ReLU). Add a Dropout layer to prevent overfitting and improve the model's generalization ability. Train the GRU network using the training set, update the network weights using the backpropagation algorithm, monitor the model performance on the validation set, use early stopping to prevent overfitting, and adjust hyperparameters such as the learning rate, batch size, and number of training epochs to optimize model performance.

[0077] S323, Input the energy consumption data of the current control cycle into the trained GRU network to obtain the energy consumption data prediction result for a future period of time;

[0078] S324, dynamically adjust the control and regulation commands based on the predicted energy consumption data.

[0079] refer to Figure 3 This application uses a trained GRU model to dynamically model time-series energy consumption data. The model takes the current energy consumption data and related features as input and outputs the predicted energy consumption value for a future period of time. Based on the prediction results, the vehicle control strategy, including power allocation and energy management, is dynamically adjusted.

[0080] In one specific embodiment of this application, S330 includes:

[0081] S331, acquire environmental image data from the environmental dynamic information, the environmental image data including: road image in front of the vehicle, traffic signs and pedestrian and vehicle images;

[0082] This application has already used the vehicle-mounted camera and V2X communication module to collect real-time images of the road ahead of the vehicle, traffic signs, and dynamic information of pedestrians and vehicles, and preprocessed them.

[0083] S332, preprocess the environmental image data to obtain preprocessed environmental image data;

[0084] The preprocessing in this step includes denoising, contrast enhancement, resizing, and data augmentation. Denoising employs a non-local means denoising algorithm to effectively eliminate image noise; contrast enhancement uses histogram equalization to improve image detail; resizing involves uniformly scaling the images to 224×224 pixels to meet CNN input requirements; and data augmentation expands the dataset through random rotation (-5° to +5°), translation (±10%), and brightness adjustment (±20%).

[0085] S333, the environmental image data and energy consumption data prediction results are respectively input into the convolutional neural network (CNN) for feature extraction, and the rolling prediction model of driving range is used to generate the rolling prediction value of driving range and the energy consumption distribution heat map for a period of time in the future.

[0086] This application constructs a CNN feature extraction network and uses an improved MobileNetV3 architecture as the backbone network to obtain a convolutional neural network (CNN). The backbone network includes depthwise separable convolutional layers to reduce the number of parameters to 3.5M and improve inference speed by 40%. An attention mechanism module is used to embed an SE attention module after the C5 feature layer to enhance the response of key features. Multi-scale feature fusion is used to concatenate the C3 and C5 feature maps after bilinear interpolation upsampling to obtain a 256-dimensional global feature vector. Multimodal feature fusion is used to fuse the visual feature vector extracted by the CNN with the temporal feature vector output by the GRU. Feature alignment is used to ensure the temporal synchronization of visual features and temporal features through timestamp matching. Attention fusion is used to calculate the similarity matrix between visual features and temporal features using a cross-attention mechanism to generate a weighted fused feature. Feature encoding is used to input the fused 512-dimensional feature vector into a fully connected layer to reduce the dimensionality to 128 dimensions as the model input.

[0087] The rolling prediction model for driving range adopts a Transformer-based prediction model, which includes an encoder for a 6-layer self-attention module to capture long-term dependencies of multimodal features; a decoder for a 3-layer cross-attention module to generate a driving range prediction for the next 120 seconds by combining historical sequences; and an output layer for a fully connected layer to output the driving range prediction values ​​for the next 10, 30, 60, and 120 seconds. A rolling prediction mechanism is introduced, updating the prediction every 1 second. The sliding window contains historical data from the past 60 seconds, and the prediction results are smoothed using an exponentially weighted moving average (EWMA) with a decay coefficient α=0.3.

[0088] Based on the spatiotemporal characteristics output by the rolling range prediction model, an energy consumption distribution map is constructed. The road 500 meters ahead of the vehicle is divided into 10m×10m grids. The predicted energy consumption is allocated to each grid cell using an inverse distance weighted interpolation algorithm. Jet color levels are used to map the energy consumption value to color intensity, generating a heat map with a resolution of 640×480. The heat map is refreshed every 5 seconds to reflect the latest prediction results.

[0089] This application employs a multi-task loss function comprising a range prediction loss and a heatmap generation loss. The range prediction loss is the Huber loss (δ=1.0), used to balance outlier robustness; the heatmap generation loss is the Structural Similarity Loss (SSIM), used to improve spatial detail representation. The training strategy first freezes the CNN parameters to train the Transformer module, then performs joint fine-tuning. A cosine annealing strategy is used with an initial learning rate of 1e-4 and a minimum learning rate of 1e-6. DropPath (p=0.1) is added between Transformer layers to prevent overfitting. INT8 quantization is used to compress the model size to 1 / 4 of its original size. The Tensor Core of the automotive NVIDIA DRIVE AGX Orin chip provides 45 TOPS of computing power. The TensorRT acceleration library is used to keep the single-frame inference time within 12ms.

[0090] In one specific embodiment of this application, the user interaction feedback layer is specifically configured as follows:

[0091] S410, in the current control cycle, displays the rolling prediction value of the driving range and the heat map of energy consumption distribution through a three-dimensional visualization interface;

[0092] S420, construct a dynamic road topology map based on the real-time traffic data and vehicle status in the environmental dynamic data;

[0093] S430: Load an intelligent route recommendation algorithm on the vehicle edge computing node, and use the intelligent route recommendation algorithm to output multiple candidate routes after inputting real-time traffic data; perform energy consumption optimization on each candidate route and select three priority routes from them;

[0094] This application presents an improved hybrid A* algorithm for loading vehicle-mounted edge computing nodes as an intelligent path recommendation algorithm. It integrates three dimensions: time cost, energy cost, and smoothness cost, and dynamically adjusts the priority through weight coefficients. =0.6 (time cost) =0.3 (energy cost) =0.1 (Ride comfort cost). The cost function is dynamically adjusted using data such as TCI; time / energy consumption / ride comfort weight coefficients, safety margin thresholds, and vehicle physical limitation parameters are set; a semi-empirical model is used, considering vehicle speed v, acceleration a, and road slope θ parameters, to calculate instantaneous energy consumption. Set minimum battery life safety margin ( (≥1.2 × estimated journey) to prevent hydrogen depletion en route. This application uses real-time traffic data as input and maps it to a cost function correction term, including congestion penalty: for road sections with TCI>0.8, the time cost weight ω_time dynamically increases by 30%; accident avoidance strategy: if an accident point is detected within 500 meters ahead, local path replanning is triggered, with a yaw angle limit of ≤15°; construction section handling: the construction area is expanded into a 50-meter buffer zone, and detour routes are forcibly planned. A dynamic road grid is generated based on real-time traffic data and vehicle status, with each grid cell labeled with the travel cost; multiple candidate paths are generated using an improved hybrid A* algorithm; energy consumption-time-smoothness multi-objective optimization is performed on the candidate paths in the rolling time domain, calculating the instantaneous energy consumption and total energy consumption of each path; a non-dominated solution set is selected in the time-energy consumption two-dimensional space, and the final path is selected according to user preference; real-time traffic data changes are monitored, and when a change in the status of a key road section is detected, local replanning is performed on the affected road section, generating new candidate paths and re-evaluating them.

[0095] refer to Figure 6 This application is based on Dynamically adjust the path search range when When the percentage is >30%, a maximum route length of 200 kilometers is allowed; when When the battery pack temperature is less than 15%, the path length is limited to ≤80 km, and an emergency hydrogen refueling station search is triggered; considering battery pack temperature limitations, when... When the temperature is above 45℃, continuous uphill sections (slope > 8%) are prohibited from being planned. A candidate path set (up to 5 paths) is generated using the hybrid A* algorithm. Model predictive control (MPC) is applied to each candidate path for 10 seconds of time-domain energy consumption optimization. Non-dominated solutions are selected in the time-energy two-dimensional space, and 3 optimal paths are retained for selection.

[0096] S440, display the three priority routes, receive feedback signals generated by the user based on the displayed information, and send them back to the multi-source heterogeneous data perception layer to correct the driving style data.

[0097] In one specific embodiment of this application, S430 includes:

[0098] S431, Load an intelligent route recommendation algorithm on the vehicle edge computing node, and use the intelligent route recommendation algorithm to output multiple candidate routes after inputting real-time traffic data; perform energy consumption optimization on each candidate route and select three priority routes from them;

[0099] S432, if the real-time traffic data shows a TCI change greater than 0.3 or a new accident report is received, route replanning is triggered, and the process re-enters S431 to perform local replanning of the TCI-changed road segment and the new accident road segment to obtain three priority routes.

[0100] This application automatically triggers route replanning every 60 seconds when a TCI change greater than 0.3 is detected within 2 kilometers ahead, or when a new accident report is received. Replanning is only performed on the affected road segments, reducing computational load. Route options are displayed through a 3D visualization interface, including a bar chart comparing time and energy consumption, energy consumption heatmaps for key road segments (e.g., red high-energy-consumption uphill sections), and hydrogen refueling station location indicators (when...). When the power consumption is less than 20%; after the driver selects a route, the feedback signal is sent back to the dynamic response optimization layer to adjust the power control strategy (such as reserving a 10% power margin to cope with sudden acceleration demands).

[0101] Secondly, this application provides a method for predicting the driving range of a fuel cell vehicle based on historical real-time energy consumption, which is implemented using the fuel cell vehicle driving range prediction system based on historical real-time energy consumption described in the first aspect.

[0102] It is worth noting that the terms "first" and "second" in this invention are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0103] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A fuel cell vehicle range prediction system based on historical real-time energy consumption, characterized by, It includes a multi-source heterogeneous data sensing layer, a dynamic response optimization layer, an adaptive prediction model layer, and a user interaction feedback layer to form a closed-loop control; The multi-source heterogeneous data perception layer is configured to collect vehicle physical parameters of the fuel cell, dynamic information of the environment around the vehicle, and driving style data in the current control cycle, and combine the three into a multimodal data pool; use the deployed vehicle edge computing nodes to preprocess the raw data in the multimodal data pool to obtain multimodal feature data, and output the driving style determination result based on the loaded hidden Markov model. The dynamic response optimization layer is configured to dynamically adjust the output power of the fuel cell stack and BOP component based on the multimodal feature data and driving style determination results in the current control cycle through a model predictive control algorithm, and combine the vehicle edge computing node to make localized real-time decisions to generate control adjustment commands and corresponding system status data. The adaptive prediction model layer is configured to aggregate model parameters from the vehicle and cloud through a federated learning framework during the current control cycle, and dynamically predict energy consumption data containing the control adjustment commands and system state data by combining a gated recurrent unit (GRU) network, so as to adjust the control adjustment commands; based on the dynamic prediction results of energy consumption data and environmental dynamic information, and by using a convolutional neural network (CNN) and a rolling prediction model for driving range, a rolling prediction value for driving range and a corresponding energy consumption distribution heatmap for a future period of time are generated. The user interaction feedback layer is configured to display the rolling prediction value of the driving range and the energy consumption distribution heat map through a three-dimensional visualization interface during the current control cycle, and to display the planned priority path based on real-time traffic data and vehicle status dynamic path planning using an intelligent path recommendation algorithm. The user's feedback signal based on the displayed information is sent back to the multi-source heterogeneous data perception layer to correct the driving style data, and then sent back to the dynamic response optimization layer to readjust the output power control. The dynamic response optimization layer is specifically configured as follows: S210, Establish a dynamic physical model of the fuel cell stack and BOP assembly; S220, using the unscented Kalman filter algorithm, the key state variables in the dynamic physical model are estimated in real time to obtain the estimated values ​​of the key state variables; S230, Discretize the dynamic physical model to obtain the discretized state space equation, and construct a rolling prediction model based on the discretized state equation; S240 determines the required power based on the driving style determination result and constructs a constrained objective function based on the required power; S250, in the current control cycle, the rolling prediction model is loaded using the vehicle-mounted edge computing node, taking the required power and the multimodal feature data as inputs, and the objective function as the objective of the rolling prediction model, so that it outputs the predicted value of the stack output voltage and the predicted value of the BOP component power consumption in the future time domain. S260, Based on the predicted output voltage of the fuel cell stack and the predicted power consumption of the BOP component, control adjustment commands and driving strategy suggestions are generated; S270, the control adjustment commands, driving strategy suggestions, and corresponding vehicle system status data are sent to the corresponding BOP component actuator via the CAN bus; wherein, the on-board edge computing node sends the first element of the control adjustment command sequence to the actuator via the CAN bus; in the next control cycle, the latest measurement values ​​are used to perform feedback correction on the rolling prediction model to correct the parameter deviation of the rolling prediction model; the first element of the control adjustment command sequence includes: air compressor speed command, hydrogen circulation pump flow command, and fuel cell stack operating temperature setpoint; the driving strategy suggestions include accelerator pedal opening and steering angle; The adaptive prediction model layer is specifically configured as follows: S310: Using local data from each vehicle, a local model is trained on the vehicle to obtain local model update parameters. The local model update parameters are then encrypted and transmitted to the cloud via a secure communication protocol. All local model update parameters are aggregated on the cloud to train a global model to obtain a new global model. The new global model update is then distributed to the vehicle via a secure communication protocol. The new global model is then retrained on the vehicle using local data. S320: The gated recurrent unit (GRU) network is used to dynamically predict the time-series energy consumption data containing the control and regulation commands and system status data, so as to adjust the control and regulation commands. The S330, based on dynamic modeling results and environmental dynamic information, generates a rolling prediction value of the driving range and the corresponding energy consumption distribution heat map for a future period of time by using a convolutional neural network (CNN) and a rolling prediction model of the driving range.

2. The fuel cell vehicle range prediction system based on historical real-time energy consumption of claim 1, wherein, The multi-source heterogeneous data sensing layer is specifically configured as follows: S110, during the current control cycle, collects fuel cell voltage, current, hydrogen pressure, temperature, and battery pack temperature in real time as vehicle physical data; S120, during the current control cycle, acquires dynamic environmental data such as road slope, rainfall, road surface friction coefficient, road images ahead of vehicles, traffic signs, and pedestrian and vehicle dynamic information through cameras, radar, and V2X communication. S130, in the current control cycle, driving style data captured by the steering wheel torque, accelerator pedal opening, and high-precision positioning module; wherein, the driving style data includes steering wheel torque signal, accelerator pedal opening data, and vehicle trajectory data; S140, the vehicle physical data, the dynamic environment data, and the driving style data are combined into a multimodal data pool; S150, the vehicle-mounted edge computing node is used to clean, standardize, and extract features from the raw data in the multimodal data pool to obtain multimodal feature data; S160, the hidden Markov model is loaded using the vehicle edge computing node, and the feature data corresponding to the driving style data is used as input to obtain the output driving style determination result.

3. The fuel cell vehicle range prediction system based on historical real-time energy consumption according to claim 2, characterized in that, S160 includes: S161, Perform FFT transformation on the steering wheel torque feature in the multimodal feature data to extract the energy proportion of the predetermined frequency band; S162, calculates the standard deviation of the rate of change of the accelerator pedal opening characteristic per unit time; S163, Based on the frequency band energy ratio and the standard deviation of the rate of change, a driving style feature vector containing the probabilities of aggressive, normal, and conservative driving styles is generated through a hidden Markov model. S164, Select the driving style with the highest probability from the driving style feature vector as the predicted driving style, and use it and the corresponding probability as the driving style determination result.

4. The fuel cell vehicle range prediction system based on historical real-time energy consumption according to claim 1, characterized in that, S320 includes: S321 collects historical energy consumption data of vehicles, preprocesses it, and then divides it into training set, validation set, and test set; S322, the pre-built GRU network is trained using the training set and validated using the trained GRU network using the validation set, so as to readjust the network parameters to obtain the trained GRU network. S323, Input the energy consumption data of the current control cycle into the trained GRU network to obtain the energy consumption data prediction result for a future period of time; S324, dynamically adjust the control and regulation commands based on the energy consumption data prediction results.

5. The fuel cell vehicle range prediction system based on historical real-time energy consumption according to claim 1, characterized in that, The S330 includes: S331, acquire environmental image data from the environmental dynamic information, the environmental image data including: road image in front of the vehicle, traffic signs and pedestrian and vehicle images; S332, preprocess the environmental image data to obtain preprocessed environmental image data; S333, the environmental image data and energy consumption data prediction results are respectively input into the convolutional neural network (CNN) for feature extraction, and the rolling prediction model of driving range is used to generate the rolling prediction value of driving range and the energy consumption distribution heat map for a period of time in the future.

6. The fuel cell vehicle range prediction system based on historical real-time energy consumption according to claim 1, characterized in that, The user interaction feedback layer is specifically configured as follows: S410, in the current control cycle, displays the rolling prediction value of the driving range and the heat map of energy consumption distribution through a three-dimensional visualization interface; S420, construct a dynamic road topology map based on the real-time traffic data and vehicle status in the dynamic environment data; S430, load an intelligent route recommendation algorithm on the vehicle edge computing node, and use the intelligent route recommendation algorithm to output multiple candidate routes after inputting real-time traffic data; Energy consumption optimization is performed on each candidate route, and three priority routes are selected. S440, display the three priority routes, receive feedback signals generated by the user based on the displayed information, and send them back to the multi-source heterogeneous data perception layer to correct the driving style data.

7. The fuel cell vehicle range prediction system based on historical real-time energy consumption according to claim 6, characterized in that, The S430 includes: S431, Load an intelligent route recommendation algorithm on the vehicle edge computing node, and use the intelligent route recommendation algorithm to output multiple candidate routes after inputting real-time traffic data; perform energy consumption optimization on each candidate route and select three priority routes from them; S432, if the real-time traffic data shows a TCI change greater than 0.3 or a new accident report is received, route replanning is triggered, and the process re-enters S431 to perform local replanning of the TCI-changed road segment and the new accident road segment to obtain three priority routes.

8. A method for predicting the driving range of a fuel cell vehicle based on historical real-time energy consumption, characterized in that, This is achieved using the fuel cell vehicle range prediction system based on historical real-time energy consumption as described in any one of claims 1 to 7.

Images