Predictive Dynamic Rule-Based Energy Management Approach for Hybrid Vehicles

Abstract

This invention discloses a predictive dynamic rule-based energy management method for hybrid electric vehicles, comprising: identifying the current operating condition of the vehicle and predicting future operating conditions on the vehicle's side, outputting future operating condition prediction features; downloading multiple sets of recommended threshold values ​​for energy management strategy rules that conform to the future operating condition prediction features from a cloud-based energy management strategy rule threshold library; ranking the recommended threshold values ​​according to the performance optimization goals of the current vehicle operation based on short-time domain vehicle speed prediction results; selecting the currently ranked top recommended threshold value as the energy management strategy execution threshold value, and switching between different operating modes of the vehicle based on the execution threshold value and state variables to form mode switching conditions. This invention can perform multi-objective energy management with comprehensive energy consumption, battery life, and vehicle emissions as objectives; it can adjust the energy management strategy rule threshold values ​​according to different performance optimization goals at different times, thereby realizing dynamic energy management of hybrid vehicles.

Description

Technical Field

[0001] This invention relates to the field of hybrid vehicle management technology, and in particular to a predictive dynamic rule-based energy management method for hybrid vehicles. Background Technology

[0002] With the increasing severity of global energy shortages and environmental pollution, the automotive industry is facing a major transformation from traditional gasoline-powered vehicles to new energy and intelligent technologies. Hybrid electric vehicles (HEVs), combining the advantages of both internal combustion engines and electric drive systems, have shown great potential in fuel economy, power performance, and emission control, and have become an important direction for the development of new energy commercial vehicles. Developing efficient and intelligent energy management strategies for hybrid commercial vehicles is not only key to improving vehicle energy consumption optimization capabilities, but also an important measure to promote the green and low-carbon transformation of the commercial vehicle industry.

[0003] Energy management strategy (EMS) for hybrid electric vehicles is a core technology for coordinating the energy distribution among the engine, motor, and battery. Its goal is to improve fuel economy, extend battery life, and optimize overall performance. Currently, mainstream strategies can be divided into two categories: rule-based energy management strategies and optimization-based energy management strategies. Rule-based energy management strategies rely on expert experience or theoretical knowledge to formulate control rules, such as thermostat control, power follower control, and optimal operating curve strategies. These strategies are logically simple and easy to implement, and are the most widely used strategies in real-world applications. However, rule-based energy management strategies rely on manual parameter tuning and cannot flexibly adapt to complex road conditions and changing driving habits. Optimization-based energy management strategies are further divided into global optimization and instantaneous optimization. Global optimization solves for the optimal energy flow based on complete operating condition information. For example, dynamic programming algorithms can achieve global optimization, but the computational load is huge, and the computational cost is too high, resulting in insufficient real-time performance. Instantaneous optimization is represented by strategies for minimizing equivalent fuel consumption and model predictive control. The strategy for minimizing equivalent fuel consumption converts electrical energy into fuel consumption through an equivalent factor, but the equivalent factor is sensitive to operating conditions and needs to be adjusted in real time. Model predictive control achieves local optima through rolling optimization and feedback correction, theoretically possessing excellent road adaptability, but requires accurate predictive models and powerful real-time computing power. Meanwhile, with increasingly stringent national emission regulations and strong consumer focus on the lifespan of hybrid systems, hybrid power strategies that only focus on single energy consumption targets can no longer meet the demands of environmental protection and market sales; multi-objective system optimization needs to be achieved in energy management.

[0004] Based on the above issues, it is clear that currently, leveraging historical experience and real-time information to achieve real-time and dynamic adjustments to energy management strategies for hybrid vehicles, and realizing collaborative adaptive matching optimization of multiple objectives such as energy consumption, emissions, power performance, and system lifespan under road conditions, presents favorable conditions. However, under current infrastructure conditions, limited computing power and network communication capabilities are significant limitations to achieving dynamic adjustments to hybrid energy management strategies using historical experience and real-time information at low cost, and are also the main challenge for the market application of predictive energy management strategies. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and defects of existing technologies by providing a predictive dynamic rule-based energy management method for hybrid electric vehicles based on the fusion of long- and short-term time-domain optimization using prior and real-time information. This method can, at the vehicle end, combine short-term vehicle speed prediction results to compare the effectiveness of recommended thresholds for different energy management strategies and dynamically select the optimal threshold based on the priority of current performance optimization goals, thereby enabling the switching of different vehicle operating modes. It can also dynamically adjust the energy management strategy of the hybrid electric vehicle based on long-term operating condition prediction information and short-term vehicle speed prediction information, thus allowing the hybrid vehicle to adapt to changes in road conditions.

[0006] A predictive dynamic rule-based energy management method for hybrid electric vehicles includes:

[0007] The vehicle-mounted terminal identifies the current operating conditions of the vehicle based on the prediction model, predicts the future operating conditions, and outputs the predicted features of the future operating conditions.

[0008] The vehicle-mounted device downloads recommended thresholds from the cloud-based energy management strategy rule threshold library, which are consistent with the predicted characteristics of future operating conditions.

[0009] The vehicle-mounted terminal combines the short-time domain vehicle speed prediction results from the short-time domain vehicle speed prediction model and sorts the recommended thresholds according to the performance optimization target priority of the current vehicle driving.

[0010] The vehicle terminal selects the currently ranked first recommended threshold as the execution threshold for the energy management strategy, and switches between different operating modes of the vehicle based on the mode switching conditions formed by the combination of the execution threshold and state variables.

[0011] The future operating condition prediction features include a future operating condition prediction curve, which is predicted by a prediction model deployed on the vehicle and uploaded to the cloud-based energy management strategy rule threshold library after the prediction is completed; the prediction model includes a repeating route prediction model and a non-repeating route prediction model.

[0012] The prediction model is a hybrid model that integrates Monte Carlo Markov Chain and LSTM. During prediction, the LSTM model first uses real-time road time-series data of a preset length as input, combined with the vehicle's historical driving sequences in the established repeating / non-repeating route road condition database, to output the probability prediction distribution of candidate sequences for the current working condition state corresponding to different working condition categories. Then, the Metropolis-Hastings algorithm is used to sample Markov chains, and the most likely working condition transition path is sampled based on the acceptance probability function constructed based on the prediction error. Through continuous sampling and acceptance judgment, the vehicle's current working condition hidden state and its transition probability matrix are dynamically updated to predict future working conditions, thus achieving the prediction of future working conditions.

[0013] The cloud-based energy management strategy rule threshold library stores recommended thresholds for a series of adaptive optimization energy management strategy rules that can be updated under various operating conditions. These recommended thresholds, combined with vehicle state variables, constitute the mode switching conditions between different vehicle operating modes. Based on the constructed energy management strategy object model, the cloud-based energy management strategy rule threshold library uses a particle swarm optimization algorithm improved from sine and cosine algorithms and flight algorithms to find the optimal recommended thresholds for different operating conditions in the constructed repeating / non-repeating route road condition library, according to vehicle parameters and their power source parameters. The recommended thresholds are different for different road conditions.

[0014] The vehicle parameters and their power source parameters include:

[0015] Vehicle parameters, including vehicle weight, vehicle frontal area, vehicle dimensions, vehicle drag coefficient, and vehicle tire size;

[0016] Engine parameters, including engine type, engine displacement, engine fuel efficiency diagram, and engine universal characteristic diagram;

[0017] Motor parameters, including full-load motor diagram and motor efficiency diagram;

[0018] Battery parameters, including battery pack structure, battery pack capacity, open circuit voltage, maximum and minimum output voltage, and internal resistance;

[0019] Transmission system parameters, including transmission ratio and transmission efficiency.

[0020] The short-time-domain vehicle speed prediction result includes a vehicle speed prediction curve within a prediction period in seconds. This vehicle speed prediction curve is formed based on the actual vehicle speed value after being transformed by the inverse normalization processing of the predicted vehicle speed prediction sequence. The actual vehicle speed value is obtained by the short-time-domain vehicle speed prediction model by processing the input vehicle speed, acceleration, and pedal opening parameters.

[0021] The short-time-domain vehicle speed prediction model is trained using a gated recurrent unit neural network (GRU).

[0022] The performance optimization targets for the current vehicle operation include battery life and / or overall energy consumption and / or vehicle emissions.

[0023] The different operating modes of the vehicle include shutdown mode, parking and charging mode, single motor drive mode, dual motor drive mode, hybrid mode with motor as the main energy source, hybrid mode with motor and engine driving together, hybrid mode with engine as the main energy source driving the vehicle and charging, braking energy recovery mode, and mechanical braking mode.

[0024] The recommended thresholds include the lower limit of remaining battery charge (SOC_L), the upper limit of remaining battery charge (SOC_H), the value where the remaining battery charge is insufficient to drive the vehicle (SOC_end), the maximum power value of the motor (Pm_max), the maximum power value in pure electric mode (Pev_max), the maximum power value of the engine (Peng_max), the vehicle speed (Vd), and the battery life (SOH_limit). The state variables include the remaining battery charge (SOC), the required torque of the vehicle (Trq_req), the required power of the vehicle (P_req), the vehicle speed, the battery health status (SOH), and the vehicle pedal opening (Brake_pedal).

[0025] This invention discloses a predictive dynamic rule-based energy management method for hybrid electric vehicles. It employs a cloud-vehicle collaborative approach for energy management, establishing an offline energy management strategy rule threshold library in the cloud, which is continuously updated as data volume increases. On the vehicle side, predicted road condition information is uploaded to the cloud, and corresponding energy management strategy thresholds are downloaded from the cloud. The effectiveness of these thresholds is tested, and the optimal threshold is applied to the on-board energy management strategy, enabling switching between different vehicle operating modes.

[0026] This invention employs a multi-model fusion architecture for both long-term and short-term predictions. The long-term operating condition prediction uses a hybrid MCMC-LSTM model, dynamically updating the state transition matrix via a Markov chain and combining it with LSTM to capture temporal dependencies for operating condition prediction. The short-term vehicle speed prediction model uses a GRU network to process vehicle speed, acceleration, and other data in real time to predict future vehicle speed information. By predicting future operating conditions and vehicle speed information, this provides a basis for the vehicle-mounted terminal to download recommended thresholds from the cloud and select energy management strategy rules, thus improving the operating condition adaptability of the dynamic rule-based energy management strategy. Attached Figure Description

[0027] Figure 1 This is a flowchart of the predictive dynamic rule-based energy management method for hybrid electric vehicles according to the present invention.

[0028] Figure 2 This is a flowchart of the present invention, which uses a particle swarm optimization algorithm (PSOSCALF) improved based on sine and cosine algorithms and flight algorithms to optimize the threshold of energy management strategy rules. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0030] The predictive dynamic rule-based energy management method for hybrid electric vehicles described in the exemplary embodiments of this application is a dynamic energy management method for hybrid electric vehicles that integrates long-term and short-term optimization of prior and real-time information. When the vehicle is driving online, it identifies the current driving condition and predicts the future driving condition based on a road condition database composed of offline long-term driving condition clustering, and selects the recommended threshold of the corresponding energy management strategy based on the driving condition prediction results. Then, it combines the vehicle speed prediction information in the future short-term domain, compares the vehicle's comprehensive energy consumption, battery life consumption, and vehicle exhaust results under the guidance of the recommended threshold of each set of energy management strategy rules, sorts the recommended values ​​according to the priority of the current performance optimization objectives, selects the threshold ranked first as the optimal energy management strategy rule threshold, and applies it to the energy management strategy at the vehicle end. By solving for the optimal threshold of the rule-based energy management strategy, the energy management control of the vehicle is realized.

[0031] See Figure 1 As shown in the exemplary embodiment of this application, the predictive dynamic rule energy management method for hybrid electric vehicles includes the following steps:

[0032] S1. The vehicle-mounted terminal identifies the current operating conditions of the vehicle based on the prediction model, predicts the future operating conditions, and outputs the predicted features of the future operating conditions.

[0033] S2. The vehicle terminal downloads recommended thresholds from the cloud-based energy management strategy rule threshold library, which are consistent with the predicted characteristics of future operating conditions;

[0034] S3. The vehicle-mounted terminal combines the short-time domain vehicle speed prediction results from the short-time domain vehicle speed prediction model and sorts the recommended thresholds according to the performance optimization target priority of the current vehicle driving.

[0035] S4. The vehicle terminal selects the currently ranked first recommended threshold as the execution threshold of the energy management strategy, and switches between different working modes of the vehicle according to the mode switching conditions formed by the combination of the execution threshold and state variables.

[0036] In the embodiments of this application, the vehicle-mounted terminal, based on the prediction model and using a road condition database composed of offline long-term operating condition clustering, identifies the current operating condition of the vehicle and predicts the future operating condition. After obtaining the predicted operating condition information, the predicted operating condition information is transmitted to the energy management strategy rule threshold database in the cloud, and multiple recommended thresholds of energy management strategy rules corresponding to the operating condition are downloaded. The vehicle-mounted terminal then selects the optimal strategy threshold by combining short-term vehicle speed prediction information. Preferably, one road / driving condition corresponds to one recommended threshold of energy management strategy rule.

[0037] In the embodiments of this application, after downloading recommended thresholds for energy management strategy rules that meet the operating conditions from the cloud based on long-term operating condition prediction results, the downloaded thresholds are combined with short-term vehicle speed prediction results and applied multiple times to the energy management strategy on the vehicle side. The overall energy consumption, battery life consumption, and vehicle exhaust results of the vehicle under the recommended thresholds are tested. The test results are compared according to the priority of the current performance optimization goals, and the downloaded thresholds are sorted based on the test results. Finally, the optimal energy management strategy rule threshold is obtained according to the sorting results, and this threshold is used as the execution threshold of the vehicle energy management strategy.

[0038] In the embodiments of this application, the future operating condition prediction feature includes a future operating condition prediction curve, which is predicted by a prediction model deployed on the vehicle. After the prediction is completed, the prediction model is uploaded to the cloud energy management strategy rule threshold library for storage, which can realize the updating and accumulation of the cloud road operating condition library. The prediction model includes a repeating route prediction model and a non-repeating route prediction model.

[0039] In this embodiment, the vehicle-mounted terminal downloads recommended thresholds for multiple sets of energy management strategy rules from the cloud-based energy management strategy rule threshold library, based on the future operating condition information predicted by the general road driving prediction model for non-repeating routes or the driving prediction model for repeating routes, and combines this with the current vehicle driving optimization objectives. These recommended thresholds are then sent to the vehicle-mounted terminal. The vehicle-mounted terminal then performs prediction calculations on the operating effect of the recommended thresholds for multiple sets of energy management strategy rules sent from the cloud, based on the vehicle speed prediction curve given by the short-time domain vehicle speed prediction model. The terminal selects the energy management strategy rule threshold that best meets the requirements and applies it to the vehicle's energy management strategy. This enables the application of energy management control in hybrid vehicles to adjust the power distribution of different power sources.

[0040] In the embodiments of this application, when the vehicle-mounted terminal predicts and calculates the operating effect of the recommended thresholds of multiple sets of energy management strategy rules sent from the cloud based on the vehicle speed prediction curve given by the short-time domain vehicle speed prediction model, it can be based on optimization targets such as battery life, comprehensive energy consumption, and vehicle exhaust emissions. According to the short-time vehicle speed prediction curve of the vehicle-mounted terminal combined with the vehicle and power system model, a rapid operation test is carried out on the vehicle-mounted terminal computing platform to obtain indicators such as economy and power. Then, according to the priority of the current performance optimization target, the performance ranking of the recommended values ​​of the energy management strategy rules sent from the cloud is obtained, and the threshold of the energy management strategy rule with the first ranking is selected as the execution threshold of the energy management strategy, which is applied to the energy management control of hybrid vehicles to adjust the power distribution of different power sources in the hybrid power system.

[0041] In the embodiments of this application, when the on-board computing platform performs a test on the downloaded recommended thresholds, the short-term vehicle speed prediction information is converted into the vehicle power demand in the future short-term domain. The vehicle power demand is then transmitted to the energy management strategy module on the vehicle. The energy management strategy module uses multiple sets of energy management strategy rule thresholds downloaded from the cloud. Under the guidance of the recommended thresholds of each set of energy management strategy rules, it determines the future working mode of the vehicle based on the power demand and vehicle status. Under this mode, with battery life, comprehensive energy consumption, and other optimization objectives, it calculates the required power output of different power sources with economic and power performance indicators. Finally, the recommended values ​​are sorted according to the priority of the current performance optimization objectives, and the recommended threshold ranked first is selected as the execution threshold or mode switching threshold.

[0042] In the embodiments of this application, the prediction model uses a hybrid model that integrates Monte Carlo Markov Chain (MCMC) and LSTM for vehicle driving condition prediction. The LSTM model outputs the probability prediction distribution of candidate sequences of driving conditions corresponding to different driving condition categories at the current time based on the historical driving sequences of vehicles in the established repeating / non-repeating route road condition database. The Metropolis-Hastings algorithm is used to sample Markov chains, and the most likely driving condition transition path is sampled based on the acceptance probability function constructed based on the prediction error. Through continuous sampling and acceptance decision, the current hidden state of the vehicle's driving condition and its transition probability matrix are dynamically updated to achieve robust prediction of future driving conditions.

[0043] In the embodiments of this application, the prediction model is trained using offline long-term historical road data. During training, the input data is offline long-term driving condition data, which is the historical road data of the target vehicle. By performing in-depth analysis, extracting driving condition features, and performing unsupervised clustering analysis based on driving condition features on large-scale long-term historical driving data, driving condition clustering is achieved to obtain typical driving condition categories and their central features, which are used as reference labels for driving condition identification. This constructs a "time-series input - driving condition label" data pair to form a road driving condition database, which is then used to train the LSTM driving condition identification model in the prediction model.

[0044] In training the LSTM condition recognition model for the prediction model, the acquired historical road data of the vehicle first needs to be preprocessed, including linear imputation of missing values ​​and detection and handling of outliers. Then, the preprocessed data is divided into short-stroke segments, or kinematic segments. During segmentation, kinematic segments are typically sliced ​​according to the continuous running time between two idling segments. For some kinematic segments that lack an idling period for an extended period, resulting in excessively long segments, segments exceeding 300 seconds are further processed by slicing them again at 100-second intervals.

[0045] After segmenting the short-stroke segments and before clustering the vehicle's driving conditions, feature parameters are extracted from the segmented kinematic segments. These extracted feature parameters consider those that significantly affect the vehicle's state under different road conditions. The feature parameters include average speed, maximum speed, average pedal opening, average acceleration, maximum acceleration, acceleration time percentage, deceleration time percentage, idling time percentage, and frequency domain features of speed. These vehicle driving condition-related feature parameters reflect the changes in road conditions during vehicle operation.

[0046] After extracting the feature parameters, when constructing a multidimensional feature parameter vector, the numerical ranges of each feature parameter may differ by orders of magnitude due to the different metric standards they adhere to. For example, the numerical ranges of vehicle speed and acceleration differ significantly, causing feature parameters with larger magnitudes to dominate during direct clustering. Therefore, to ensure that all feature parameters reflect their influence during analysis, a unified standardization process is required for the extracted feature parameters. Given that the vehicle feature parameters collected in the experiment are complete and their extreme values ​​are easily obtained, the Min-Max standardization method can be used to perform dimensionless processing on the feature parameters.

[0047] The Min-Max standardization, also known as deviation standardization, scales the data to the interval [0,1] by subtracting the minimum value from each data point and then dividing by the difference between the maximum and minimum values, as shown in the following formula.

[0048]

[0049] In the formula, Xi represents the standardized data, and x min x is the minimum value in the data. max x represents the maximum value in the data. i This is the original data.

[0050] Based on the aforementioned characteristic parameters such as vehicle speed and average pedal opening, cluster analysis is performed on the driving conditions. The multidimensionality of the data increases the complexity of the clustering solution, and the selection of characteristic parameters is somewhat subjective, with different characteristic parameters having varying importance. To preserve the original characteristic parameter information, simplify the calculation process, and make the clustering results more accurate, this embodiment selects Principal Component Analysis (PCA) to reduce the dimensionality of the characteristic parameters. PCA, as a common data dimensionality reduction method, transforms the feature vectors into a set of uncorrelated principal components through orthogonal transformation; the specific steps are as follows:

[0051] (1) Calculate the correlation matrix R of multiple characteristic parameters such as vehicle speed and average pedal opening:

[0052]

[0053] In the formula, x ki and x kj These represent the standardized feature parameter values.

[0054] (2) Calculate eigenvalues ​​and eigenvectors

[0055] u j =(u 1j ,u 2j ,...,u ij ) T ;

[0056] In the formula, u j Representing each principal component vector, that is, the principal component vector composed of input features, is the j-th eigenvector of the correlation matrix R.

[0057] (3) Calculate the variance contribution rate and cumulative contribution rate of the principal components.

[0058]

[0059] In the formula, P j The eigenvalue λ represents the contribution rate of the j-th component. j Let represent the variance of the j-th principal component.

[0060] (4) Calculate the principal component composite score

[0061]

[0062] In the formula, y j This represents the score of the j-th principal component.

[0063] After dimensionality reduction of the principal components, spectral clustering is used to perform unsupervised clustering of driving conditions, i.e., fragment clustering, to obtain fragment clustering results, forming multiple different driving conditions for training a multi-layer LSTM driving condition recognition model. The steps of the spectral clustering algorithm are as follows:

[0064] (1) Construct a similarity matrix using the Gaussian kernel function: W ij The element in the i-th row and j-th column of a similar matrix.

[0065]

[0066] In the formula, σ represents the Gaussian kernel parameter. The Radial Basis Function (RBF) is chosen as the Gaussian kernel function for constructing the similarity matrix. This represents the squared Euclidean distance between sample i and sample j (two driving condition segments).

[0067] (2) According to W ij Calculate the Laplace matrix:

[0068] Calculate the degree matrix based on the similarity matrix:

[0069] Calculate the unnormalized Laplacian matrix from the degree matrix: L = DW;

[0070] In the formula, W is the similarity matrix, which reflects the degree of similarity between two driving conditions, and D is the degree matrix.

[0071] Based on the following, the normalized Laplace matrix is ​​obtained: L sym =D -1 / 2 LD -1 / 2 ;

[0072] (3) Calculate the eigenvectors and eigenvalues ​​of the Laplacian matrix, where x is the corresponding eigenvector;

[0073] Lx = λDx, where λ represents the regularization parameter; Lx represents the regularized Laplacian matrix components, and Dx represents the normalized degree matrix.

[0074] (4) Combine the eigenvectors of the k regularized Laplacian matrix components to form a new matrix;

[0075] (5) Use the K-means clustering method to cluster the new feature space;

[0076] (6) The clustering results return the original data, and the working condition segments are clustered into categories such as city, highway, and mountain.

[0077] Through the above processing, clustering of the segmented historical road data fragments was achieved.

[0078] Next, the segmented historical road data fragments are labeled according to the clustering results to construct a "time-series input - condition label" data pair for training the LSTM condition recognition model. This data pair is then used to train a multi-layer LSTM condition recognition model to model the temporal evolution characteristics of driving conditions, thereby obtaining an LSTM condition recognition model that can predict the current condition category distribution based on the input time series.

[0079] Through the above techniques, data collection, preprocessing, feature parameter extraction, clustering and labeling were completed, and the LSTM working condition recognition model was trained.

[0080] After the LSTM working condition recognition model is trained, when it is used for online working condition identification, it is combined with the constructed road working condition database to identify the current working condition type of the target vehicle and predict its future working condition. The predicted probability distribution of the working condition category at the current moment provided by the LSTM working condition recognition model is used as the initial probability distribution of the MCMC algorithm, and then the Metropolis-Hastings sampling algorithm is executed to predict the future working condition type. In this embodiment, the execution of working condition identification and prediction by the LSTM working condition recognition model and the MCMC algorithm specifically includes:

[0081] Vehicle operation data, including speed, throttle opening, acceleration, and other dimensions, is continuously collected using a sliding time window to form a fixed-length time-series data set. The most recent fixed-length time-series data (real-time road data) is input into a deployed LSTM (Laser-Based Detailed Module) road condition recognition model. This model, combined with a constructed road condition database, models the dynamic features of the time series through an internal memory mechanism, outputting a predicted probability distribution / initial road condition probability matrix indicating the vehicle's current position within various typical road condition categories. The predicted probability distribution reflects the most likely current road condition type the vehicle is in under the current data pattern, providing prior information for subsequent road condition trajectory sampling.

[0082] In the MCMC algorithm, the current LSTM output is first used as the initial state distribution. Since the LSTM provides the predicted probability distribution of the current work condition category, and MCMC aims to sample from a target distribution, it is used as an approximation of that target distribution. To execute the Metropolis-Hastings sampling algorithm, the proposal distribution is defined, the acceptance probability is constructed and calculated, and sample acceptance is determined; the steps are as follows:

[0083] (1) Proposal distribution formulation:

[0084] The proposal distribution is set based on the working condition state transition matrix obtained from historical data statistics, that is, from the current state C t The transition probability q(C′|C) to candidate state C′ t Because the changes in operating conditions exhibit Markov properties, the transitions in vehicle operating conditions over a short period of time show strong local regularity. Therefore, using the transition matrix to model the proposal distribution is not only reasonable, but also effectively guides the sampling process to closely reflect reality.

[0085] (2) Construct and calculate the acceptance probability;

[0086] The calculation of the acceptance probability follows the principles of the Metropolis-Hastings algorithm, and the expression is:

[0087]

[0088] Where p(C′) and p(C) t Let q represent the probability values ​​of the candidate states and the current state in the target distribution, which are the predicted probabilities of the LSTM working condition recognition model for each working condition category. Since the proposal distribution q comes from the state transition matrix, and the matrix is ​​usually symmetric or approximately symmetric, i.e., q(C′∣C) t )≈q(C t ∣C′), the proposal probability term in the above equation can be canceled out, thus simplifying the acceptance probability to:

[0089]

[0090] Through the above process, in each sampling step, the algorithm tends to accept states with higher LSTM prediction probabilities, thereby improving the algorithm's good target orientation while retaining a certain degree of randomness to achieve diversity of target samples.

[0091] (3) Sample acceptance judgment:

[0092] The acceptance judgment process is based on the acceptance probability. In each sampling step, a uniformly distributed random number u between 0 and 1 is generated. If u ≤ α, then the candidate state C′ is accepted as the update of the current working condition state; otherwise, the original state C is retained. t This process ensures that the state update process conforms to the Markov chain properties, and that the sampling sequence under long-term operation will approximate the true target distribution.

[0093] Through the above steps, starting from the current state, multiple future time step work condition evolution trajectories are constructed through MCMC path sampling, which can be used for subsequent work condition distribution analysis, strategy decision feedforward input, or rolling optimization tasks.

[0094] The training-based LSTM-based work condition prediction model in this application combines LSTM prediction with MCMC inference for long-term work conditions. It can predict future vehicle speed-acceleration trajectories, traffic conditions, or energy consumption demands over a relatively long time range (from minutes to hours, or even the entire travel task cycle) based on historical work condition data, road features, and environmental information. It can not only capture the temporal patterns in work condition changes but also reflect the uncertainty of multiple paths, thus improving the robustness and practicality of the prediction.

[0095] In this embodiment, the "time-series input - condition label" data pairs used to train the LSTM model for driving condition identification are constructed into a condition library containing all road conditions for training the prediction model. This road condition library includes a general road condition library and a feature road condition library, stored in the energy management strategy rule threshold library. The feature road condition library consists of repetitive routine route conditions, tailored to specific users and composed of predetermined routes they travel daily. It includes traffic condition features, time features, and vehicle operation features of these repetitive routes, used for training the repetitive route prediction model. The repetitive route prediction model is used to construct condition prediction curves for repetitive routes of specific user vehicles. The routes in the general road condition library are non-repetitive routes, used to train a non-repetitive route prediction model for general road driving prediction, and used for predicting general or non-repetitive road driving routes for ordinary user vehicles and constructing condition prediction curves. The general road condition library is shown in Table 1.

[0096] Table 1

[0097]

[0098] The general road conditions are divided into four main categories based on driving scenarios: urban conditions, highway conditions, mountain conditions, and suburban conditions. Each driving scenario is further subdivided into several specific driving scenarios. The characteristic road conditions consist of repetitive routine route conditions, which are tailored to specific users and consist of predetermined routes that they drive every day. These conditions include the traffic conditions, time characteristics, and vehicle operation characteristics of the route.

[0099] The urban driving conditions are further subdivided into congested urban core areas, traffic light intersections (urban road intersections), off-peak urban main roads, urban commercial areas, and urban expressways. The characteristics of congested urban core areas include: extremely low vehicle speeds and a high proportion of idling time, such as on main roads and roads surrounding commercial areas during morning and evening rush hours; frequent starts and stops, significant acceleration fluctuations, and the engine often operating at inefficient idling speeds, with the electric motor frequently participating in drive and energy recovery. Therefore, energy management strategies in this scenario require reducing engine idling and enhancing electric motor energy recovery. The characteristics of traffic light intersection scenarios are: centered around intersections, encompassing a complete process of "approaching the intersection - decelerating - stopping - starting - leaving," with a speed curve exhibiting a trapezoidal shape. Therefore, energy management strategies need to rationally optimize the timing of energy recovery before stopping. Off-peak conditions on urban arterial roads, such as continuous straight traffic without intersection interference or intersection interference where vehicles only need to slow down to pass without coming to a complete stop, are characterized by moderate speeds, smooth traffic flow, no continuous congestion, regular traffic light cycles, and minimal acceleration fluctuations. In these scenarios, the energy management strategy needs to maintain the engine's efficient operating range to balance state of charge (SOC). Urban commercial area scenarios are characterized by dense pedestrian and non-motorized vehicle traffic, speed restrictions, and frequent low-speed yielding. Therefore, the energy management strategy requires a fast-responding motor to reduce power switching delays. Urban expressways, on the other hand, are characterized by no traffic light interference, higher speeds, and the ability to maintain a constant speed for extended periods with relatively stable acceleration. Therefore, the energy management strategy requires engine-dominated drive with the motor assisting as needed to reduce fuel consumption.

[0100] The high-speed operating conditions are further subdivided into smooth-flowing high-speed conditions, heavy-traffic high-speed conditions, and high-speed entrance / exit / service area conditions. Smooth-flowing high-speed conditions are characterized by vehicle speeds remaining stable at high speeds for extended periods with minimal acceleration fluctuations. In this scenario, the energy management strategy must maintain the engine within its efficient operating range, with the electric motor only providing brief assistance during overtaking to avoid frequent power switching and reduce energy consumption. Heavy-traffic high-speed conditions require vehicles to accelerate and decelerate slightly to follow the rhythm of the vehicle in front, resulting in speed fluctuations. Therefore, the energy management strategy must enable the electric motor to participate in smoothing speed fluctuations and reducing engine load fluctuations. High-speed entrance / exit / service area conditions involve a process of "decelerating into - low-speed driving - accelerating out," thus requiring the energy management strategy to enhance electric motor energy recovery during deceleration and coordinate engine and electric motor output during acceleration.

[0101] The mountain driving conditions are further subdivided into continuous uphill driving, continuous downhill driving, and winding mountain road driving. The continuous uphill driving scenario is characterized by consistently high vehicle load output. Therefore, the energy management strategy should prioritize engine-driven operation, with the electric motor providing short-term supplementary power (to avoid engine overload), while limiting deep battery discharge. The continuous downhill driving scenario requires frequent braking, so the energy management strategy should maximize electric motor energy recovery to avoid overheating of mechanical braking, while also reducing battery charge in advance to prevent overcharging during electric motor energy recovery. The winding mountain road driving scenario requires frequent deceleration (before entering the curve) and acceleration (after exiting the curve), therefore the energy management strategy needs to be able to quickly respond to power demands, with the electric motor assisting in seamlessly transitioning between energy recovery and power output modes.

[0102] The suburban driving conditions are further subdivided into suburban road driving conditions (township main road driving conditions), suburban industrial zone driving conditions, and suburban expressway driving conditions. The suburban driving scenario is characterized by the absence of traffic lights but numerous intersections, resulting in significant fluctuations in vehicle acceleration. Therefore, the energy management strategy needs to achieve reasonable power output from the engine and motor, reducing frequent switching of drive modes. The suburban industrial zone driving scenario is characterized by a high density of large trucks, relatively flat roads but unstable traffic flow, and frequent overtaking. Therefore, the corresponding energy management strategy needs to strengthen power redundancy reserves to cope with the high power demands during acceleration and overtaking. The suburban expressway driving scenario is characterized by higher vehicle speeds and smaller acceleration fluctuations. Therefore, the energy management strategy needs to ensure the engine operates within its high-efficiency range and that the battery maintains a stable state of charge (SOC).

[0103] In this embodiment, the cloud-based energy management strategy rule threshold library includes the road condition library and energy management strategy rule thresholds that match it one by one. This allows for the dynamic selection of the corresponding energy management strategy based on the condition information. For each series of updatable conditions, an optimization algorithm is used to find the recommended threshold for the corresponding energy management strategy rule, thereby forming the cloud-based energy management strategy rule threshold library. The recommended threshold, combined with vehicle state variables, constitutes the mode switching conditions between different vehicle operating modes, thus forming the vehicle energy management strategy.

[0104] In this embodiment of the application, for different road conditions and the vehicle's own status, a prediction model can be deployed in the cloud to construct a condition prediction curve. With the resources of the cloud platform, according to the comprehensive goals of optimal energy consumption, optimal emissions, optimal power performance or optimal system life, an optimization algorithm is used to optimize the mode switching threshold of the energy management strategy, forming a series of condition-adaptive optimized energy management strategy rule thresholds. These thresholds are then summarized and stored in the cloud's energy management strategy rule threshold library and updated and iterated as data accumulates.

[0105] The cloud-based system can continuously adjust its energy management optimization objectives based on the vehicle's constantly changing state and varying road conditions. While meeting power demands, these objectives include economic indicators represented by overall energy consumption, emission regulations represented by vehicle emissions, and reliability indicators represented by battery life. Of these, overall energy consumption and emission information can be directly retrieved from the vehicle controller. Battery life, as a reliability optimization objective, requires quantification based on battery operating conditions, and the quantification result is used as one of the optimization indicators to guide the determination of energy management strategy rule thresholds.

[0106] In a preferred embodiment, the cloud-based energy management strategy rule threshold library is based on a constructed energy management strategy object model. According to vehicle parameters and their power source parameters, a particle swarm optimization algorithm improved from sine and cosine algorithms and flight algorithms is used to find recommended thresholds for optimal energy management strategy rules corresponding to different operating conditions in the constructed repetitive / non-repetitive route road condition library. For example, with the goal of minimizing overall energy consumption, the recommended thresholds for different operating conditions are found. The recommended thresholds for energy management strategy rules under different operating conditions are then aggregated to obtain the energy management strategy rule threshold library. Furthermore, the recommended thresholds differ for different road conditions, ensuring that different modes are switched based on the comparison of system efficiency across different modes, guaranteeing optimal overall energy efficiency throughout the cycle.

[0107] The method described above uses a particle swarm optimization algorithm (PSOSCALF) based on the sine and cosine algorithms and the flight algorithm to find the recommended threshold (mode switching threshold) of the optimal energy management strategy rule for different working conditions in the constructed repeating / non-repeating route road condition database. This threshold is obtained by optimizing the optimal cost function under different working conditions in the road condition database, as shown in the figure. First, initialize the population and the mode switching threshold parameter in the Particle Swarm Optimization (PSOSCALF) algorithm. Input the mode switching threshold parameter into the Simulink–Cruise M co-simulation model for calculating the fitness function to jointly simulate the parameter values ​​to be optimized (recommended threshold). Calculate the objective function value of the current particle position, and calculate the individual extreme value and the population extreme value based on the fitness function. Update the particle velocity and position by comparing the individual extreme value, the population extreme value, and the objective function value. After updating the particle velocity and position, determine whether the maximum number of iterations or the required accuracy has been reached. If so, end the optimization process; otherwise, update the mode switching threshold parameter and input it into the Simulink–Cruise M co-simulation model for calculating the fitness function to jointly simulate the parameter values ​​to be optimized. Repeat the above optimization process until the requirements are met. The particle position corresponds to the recommended threshold of the energy management strategy rule. The particle velocity and position update equations in the Particle Swarm Optimization (PSOSCALF) algorithm, which combines the sine and cosine algorithms and the flight algorithm, are as follows:

[0108] V i (t+1)=wV i (t)+c1r1(X pBest -X i (t))+c2r2(X gBest -X i (t))

[0109]

[0110] In the formula, V j (t+1) and X i (t+1) represents the velocity and position vectors of particle i at iteration t+1, respectively; c1 and c2 represent the individual learning factor and the social learning factor, respectively; w is the inertia factor; X i (t) is the particle position vector, Levy walk (X i (t) represents the particle position vector added to Levy Flight, r1, r2, r3, r4 are random numbers, and X pBest Let X be the optimal point of the individual after t iterations. gBest It is the global optimum after t iterations, r4 is the switching factor between sine and cosine, r1 specifies the next position direction of the solution, r2 defines the distance towards or away from the endpoint, and r3 introduces random weights to the endpoint.

[0111] To ensure both the improved algorithm's exploration capability, enabling the Particle Swarm Optimization (PSOSCALF) algorithm to quickly reach the optimal region, and to maximize its utilization to find the optimal value, the parameters of the sine and cosine algorithms need to be continuously adjusted based on the actual search process, with r1 gradually decreasing during optimization iterations.

[0112] This application employs a particle swarm optimization algorithm, improved from sine and cosine algorithms and the flight algorithm, to optimize the mode switching threshold for the energy management strategy corresponding to each operating condition. By finding the system efficiency-optimal mode switching threshold for different modes, i.e., the energy management strategy rule threshold, the overall energy efficiency of the energy management execution scheme is optimized.

[0113] In this embodiment of the application, the vehicle parameters and its power source parameters include:

[0114] Vehicle parameters, including vehicle weight, vehicle frontal area, vehicle dimensions, vehicle drag coefficient, and vehicle tire size;

[0115] Engine parameters, including engine type, engine displacement, engine fuel efficiency diagram, and engine universal characteristic diagram;

[0116] Motor parameters, including full-load motor diagram and motor efficiency diagram;

[0117] Battery parameters, including battery pack structure, battery pack capacity, open circuit voltage, maximum and minimum output voltage, and internal resistance;

[0118] Transmission system parameters, including transmission ratio and transmission efficiency.

[0119] In this embodiment of the application, the different operating modes of the vehicle include STOP mode, STOP_char mode, single motor drive mode (EV1), dual motor drive mode (EV2), hybrid mode with motor as the main energy source (HEV_1), hybrid mode with motor and engine jointly driven (HEV_2), hybrid mode with engine as the main energy source driving the vehicle and charging (HEV_3), brake energy recovery mode (Brake_gen), and mechanical braking mode (Brake_fri). The different operating modes of the vehicle correspond to four driving states: parking state, driving state, coasting state, and braking state.

[0120] The recommended thresholds include the lower limit of remaining battery charge (SOC_L), the upper limit of remaining battery charge (SOC_H), the value where the remaining battery charge is insufficient to drive the vehicle (SOC_end), the maximum power value of the motor (Pm_max), the maximum power value in pure electric mode (Pev_max), the maximum power value of the engine (Peng_max), the vehicle speed (Vd), and the battery life (SOH_limit). The state variables include the remaining battery charge (SOC), the required torque of the vehicle (Trq_req), the required power of the vehicle (P_req), the vehicle speed, the battery health status (SOH), and the vehicle pedal opening (Brake_pedal).

[0121] In the embodiments of this application, the energy management strategy corresponding to each road condition is built on a layered architecture, with all operation and control modes divided into a driving mode layer and an energy management execution layer. Based on the actual working state of the vehicle, the vehicle operating mode is set to multiple specific drive modes / working modes, each with a corresponding energy management execution scheme. The energy management strategies for different road conditions include multiple modes such as pure electric, hybrid, engine direct drive, and energy recovery. The main difference between the energy management strategies for different road conditions lies in the different mode switching conditions that trigger the switching between each working mode.

[0122] The mode switching conditions are shown in Table 2 below:

[0123] Table 2

[0124]

[0125] The system utilizes a cloud-based energy management strategy rule threshold library, including a road condition library and recommended thresholds for energy management strategy rules. The road condition library comprises four general driving scenario conditions and a feature road condition library consisting of repetitive routine route conditions. Each of the four general driving scenario conditions is further subdivided into several specific scenario conditions. Simultaneously, the energy management strategy formulated for each specific condition includes hierarchical control logic for four driving states and nine driving modes. Furthermore, the PSOSCALF algorithm, which integrates sine and cosine algorithms with flight algorithms, optimizes the energy management rule thresholds for different conditions in the cloud, ultimately achieving dynamic energy management based on different driving scenarios of hybrid vehicles.

[0126] The short-time-domain vehicle speed prediction result includes a vehicle speed prediction curve within a prediction period measured in seconds, which is predicted by a short-time-domain vehicle speed prediction model deployed at the vehicle end.

[0127] In the embodiments of this application, the short-time-domain vehicle speed prediction model is trained using a gated recurrent unit neural network (GRU). It processes the input vehicle speed, acceleration, and pedal opening parameters to output a future vehicle speed prediction sequence or a vehicle speed prediction value vector. The initial parameters of the short-time-domain vehicle speed prediction model are derived from offline training and development using historical experience data in the cloud. Short-time-domain vehicle speed prediction refers to a method of predicting future vehicle speeds using recent actual vehicle speed data as input within a prediction period measured in seconds. The short-time-domain prediction model runs on the vehicle and serves as the computational boundary condition for the dynamic fusion optimization of the hybrid power system's dynamic energy management strategy. For example, by inputting a matrix of vehicle speed data from the past 10 seconds into the GRU, predictions for the next 10 steps are made, resulting in a prediction of the vehicle speed sequence for the next 10 seconds. This means the GRU model outputs a vehicle speed prediction value vector containing the next 10 seconds' worth of data, and the prediction results are inversely normalized to convert them into actual vehicle speed values. In the offline training of the GRU model, the large-scale vehicle driving history data covering different driving scenarios is first preprocessed. The preprocessed data is then divided into training samples according to time windows. Each training sample contains 10 seconds of historical data and the corresponding label value for the next 10 seconds, constructing an input-output pair. This input-output pair is then used as the input to the designed GRU model for training. The Adam optimizer minimizes the mean squared error loss function. During training, an early stopping strategy and a learning rate decay mechanism are used to improve the model's generalization ability. After training, cross-validation is used to evaluate the model's performance, and the prediction accuracy of the model in different scenarios is verified using metrics such as root mean square error (RMSE) and mean absolute error (MAE).

[0128] After the short-time-domain vehicle speed prediction model is trained, to ensure efficient and stable operation and provide accurate prediction results in real time during online vehicle speed prediction, onboard sensor data, including vehicle speed, acceleration, and pedal opening parameters, can be collected at a frequency of 100Hz and stored in a sliding window buffer, maintaining a buffer data length of 10 seconds. When new data arrives, the oldest data is automatically removed, while maintaining the window size. Real-time preprocessing is performed on the buffer data, using normalization parameters saved during training to standardize the data and reshape it into a three-dimensional tensor conforming to the model input format. The trained GRU model is loaded, and the preprocessed data is input into the GRU model to obtain the vehicle speed prediction sequence for the next 10 seconds. Then, the prediction results are denormalized to convert them into actual vehicle speed values. To improve the real-time performance of the prediction, an asynchronous processing mechanism can be used, allowing the prediction process to run in parallel with data acquisition.

[0129] In the embodiments of this application, the performance optimization objectives for the current vehicle operation include battery life and / or overall energy consumption and / or vehicle emissions. In hybrid electric vehicles, the battery serves as a power source, and its lifespan assessment is a crucial consideration for vehicle energy management. This assessment can be based on the battery's State of Health (SOH) as an energy management parameter. Battery lifespan can be assessed using the charge flux method. The charge flux method is a commonly used method for estimating remaining battery lifespan. It is a semi-empirical model that fits key parameters to experimental data, estimates battery lifespan damage in real-time within a sliding time domain (short-term charge-discharge cycles), and uses the concept of a severity factor to quantify battery lifespan. Specific methods include:

[0130] In its modeling process, the battery degradation capacity Q is first defined. loss％ As shown in the following formula:

[0131] Q loss％ =σ·Ah z ;

[0132]

[0133] In the formula, Q loss％ To determine the extent of battery capacity degradation, a 20% limit was set, and σ was a severity factor used to quantify the real-time loss of battery life during HEV driving. α, β, and z were obtained by fitting battery degradation experimental data. E a The total activation energy is related to the battery discharge rate I. c Linear correlation, R is the gas constant, Ah is the charge flux, describing the total charge throughput over the battery's entire lifespan, T cell The cell temperature is calculated by establishing a Debert automotive battery pack temperature field model, based on the heat transfer between the cell layer, inner wrapping layer, and outer protective layer.

[0134] The severity factor σ can also be calculated as the ratio of the estimated total battery charge flux under rated operating conditions to the estimated charge flux under real-time operating conditions. When the real-time battery charge flux is greater than the rated operating condition charge flux, i.e., σ is greater than 1, it indicates that the battery life loss under the current operating condition is greater than that under the rated operating condition. The expression is as follows:

[0135]

[0136] In the formula, Γ is the rated charge flux, γ is the actual charge flux, EOL (End of Life) is the battery life endpoint, and I is the battery charge flux. nom I(t) represents the battery charging and discharging current under rated operating conditions, and I(t) represents the battery charging and discharging current under real-time operating conditions.

[0137] The effective charge flux Ah is obtained by integrating σ and |I(τ)|. eff Characterizes the degree of battery life degradation:

[0138]

[0139] In the formula, I(τ) represents the battery charging and discharging current in the sliding time domain. As σ and the battery charging and discharging current increase, Ah... eff The higher the value, the greater the impact of the current operating conditions on battery life. When Ah eff When the accumulated charge flux reaches the rated charge flux Γ, it indicates that the battery has reached its lifespan limit. Therefore, the battery state of health (SOH) is defined as shown in Equation 5:

[0140]

[0141] Generally, the battery capacity degradation limit is defined as 20%, that is, when Q loss When % = 20%, SOH = 0%.

[0142] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and therefore all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

[0143] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A predictive dynamic rule-based energy management method for hybrid electric vehicles, characterized in that, Includes the following steps: The vehicle-mounted terminal identifies the current operating conditions of the vehicle based on the prediction model, predicts the future operating conditions, and outputs the predicted features of the future operating conditions. The vehicle-mounted device downloads recommended thresholds from the cloud-based energy management strategy rule threshold library, which are consistent with the predicted characteristics of future operating conditions. The vehicle-mounted terminal combines the short-time domain vehicle speed prediction results from the short-time domain vehicle speed prediction model and sorts the recommended thresholds according to the performance optimization target priority of the current vehicle driving. The vehicle terminal selects the currently ranked first recommended threshold as the execution threshold for the energy management strategy, and switches between different operating modes of the vehicle based on the mode switching conditions formed by the combination of the execution threshold and state variables. The future operating condition prediction features include future operating condition prediction curves, which are predicted by prediction models deployed on the vehicle end, and uploaded to the cloud-based energy management strategy rule threshold library after the prediction is completed; the prediction models include repeating route prediction models and non-repeating route prediction models. The prediction model is a hybrid model that integrates Monte Carlo Markov Chain and LSTM. During prediction, the LSTM model first uses real-time road time-series data of a preset length as input, combined with the vehicle's historical driving sequences in the established repeating / non-repeating route road condition database, to output the probability prediction distribution of candidate sequences for the current working condition state corresponding to different working condition categories. Then, the Metropolis-Hastings algorithm is used to sample Markov chains, and the most likely working condition transition path is sampled based on the acceptance probability function constructed based on the prediction error. Through continuous sampling and acceptance judgment, the vehicle's current working condition hidden state and its transition probability matrix are dynamically updated to predict future working conditions, thus achieving the prediction of future working conditions.

2. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The cloud-based energy management strategy rule threshold library stores recommended thresholds for a series of adaptively optimized energy management strategy rules that can be updated under various operating conditions. These recommended thresholds, combined with vehicle state variables, constitute the mode switching conditions between different vehicle operating modes. Based on the constructed energy management strategy object model, the cloud-based energy management strategy rule threshold library uses a particle swarm optimization algorithm improved from sine and cosine algorithms and flight algorithms to find the optimal recommended thresholds for different operating conditions in the constructed repeated / non-repeated route road condition library, according to vehicle parameters and their power source parameters. The recommended thresholds differ for different road conditions.

3. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 2, characterized in that, The vehicle parameters and its power source parameters include: Vehicle parameters, including vehicle weight, vehicle frontal area, vehicle dimensions, vehicle drag coefficient, and vehicle tire size; Engine parameters, including engine type, engine displacement, engine fuel efficiency diagram, and engine universal characteristic diagram; Motor parameters, including full-load motor diagram and motor efficiency diagram; Battery parameters, including battery pack structure, battery pack capacity, open circuit voltage, maximum and minimum output voltage, and internal resistance; Transmission system parameters, including transmission ratio and transmission efficiency.

4. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The short-time-domain vehicle speed prediction result includes a vehicle speed prediction curve within a prediction period in seconds. This vehicle speed prediction curve is formed based on the actual vehicle speed value after being transformed by the inverse normalization processing of the predicted vehicle speed prediction sequence. The actual vehicle speed value is obtained by the short-time-domain vehicle speed prediction model by processing the input vehicle speed, acceleration, and pedal opening parameters.

5. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The short-time-domain vehicle speed prediction model is trained using a gated recurrent unit neural network (GRU).

6. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The current vehicle performance optimization targets include battery life and / or overall energy consumption and / or vehicle emissions.

7. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The different operating modes of the vehicle include shutdown mode, parking and charging mode, single motor drive mode, dual motor drive mode, hybrid mode with motor as the main energy source, hybrid mode with motor and engine driving together, hybrid mode with engine as the main energy source driving the vehicle and charging, braking energy recovery mode, and mechanical braking mode.

8. The predictive dynamic rule-based energy management method for hybrid electric vehicles according to claim 1, characterized in that, The recommended thresholds include the lower limit of remaining battery charge (SOC_L), the upper limit of remaining battery charge (SOC_H), the value where the remaining battery charge is insufficient to drive the vehicle (SOC_end), the maximum power value of the motor (Pm_max), the maximum power value in pure electric mode (Pev_max), the maximum power value of the engine (Peng_max), the vehicle speed (Vd), and the battery life (SOH_limit). The state quantities include the remaining battery charge (SOC), the required torque of the vehicle (Trq_req), the required power of the vehicle (P_req), the vehicle speed, the battery health status (SOH), and the vehicle pedal opening (Brake_pedal).

Images