Longitudinal stability control system for distributed drive electric vehicles based on multi-model MPC Authors: Meng Dang, Chuanwei Zhang, Zhongyu Guo, Jianlong Wang Published: 2026 ABSTRACT AbstractDistributed drive electric vehicles (DDEVs) show great potential in electric vehicle applications owing to their high transmission efficiency and precise driving control. Nevertheless, existing stability control methods suffer from insufficient proactivity and limited real-time regulation capability, making them unable to guarantee satisfactory longitudinal stability performance under complex operating conditions. To address this issue, this paper presents a vehicle state estimation and longitudinal stability control system for complex driving scenarios. A LiDAR-IMU fusion scheme is developed, which combines the adaptive unscented Kalman filter (AUKF) and time-series analysis (TSA) to improve state estimation accuracy. Furthermore, a multi-model model predictive control (MPC) framework is established to classify driving conditions and generate integrated control commands through weighted fusion, so as to achieve optimized longitudinal stability and smooth mode switching. The main novelty of this work lies in the integration of predictive state estimation and scenario-classified weighted-fusion multi-model MPC, which differs from conventional switching MPC and gain-scheduled MPC by avoiding abrupt mode switching and explicitly considering model differences under typical DDEV operating conditions. Both simulation and hardware-in-the-loop (HIL) results validate that the proposed system effectively enhances longitudinal stability and control performance under complex conditions. The root-mean-square error (RMSE) of yaw rate estimation is reduced to 0.111 deg/s, and the control accuracy is improved by 21.7% compared with the conventional MPC method. This work lays a solid theoretical basis for the application of distributed drive electric vehicles.Keywords: Distributed drive, Electric vehicles, Stability control, Model predictive control, Multi-model fusion, State estimationSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction Due to environment pollution, electric vehicles (EVs) represent a major trend in the future development direction as their high efficiency, environmental friendliness, and low-carbon characteristics1,2. Among these, DDEVs have attracted considerable attention, as their actuated wheel motors enable efficient power transmission and flexible torque distribution3,4. Compared to conventional centralized-drive vehicles, DDEVs eliminate mechanical transmission systems, thereby providing substantial advantages in power allocation, dynamic control, and energy efficiency management, particularly with regard to intelligent driving and adaptability to complex operating conditions5,6. For DDEVs, fully leveraging their independent torque distribution capability to improve vehicle stability and safety under diverse operating scenarios constitutes a central research focus in the field of automotive control systems7,8. Longitudinal stability constitutes a crucial indicator for assessing vehicle dynamic performance and driving safety in the context of longitudinal motion9. Under rapid acceleration, emergency braking, or complex road conditions, such as slippery surfaces or varying friction coefficients, vehicles often experience diminished longitudinal stability due to uneven torque distribution or wheel slip, potentially resulting in loss of control and other safety-critical incidents10. In this study, longitudinal stability is defined as the ability of a DDEV to maintain stable longitudinal tire-force generation and bounded wheel slip during acceleration, braking, and road-adhesion variations. Because the four wheels of a DDEV are independently driven, uneven longitudinal torque distribution may also generate additional yaw moment and sideslip response. Therefore, longitudinal stability in this paper includes not only wheel slip suppression and longitudinal velocity regulation, but also the prevention of yaw and sideslip instability induced by longitudinal force imbalance. Conventional longitudinal stability control methods, which primarily rely on mechanical braking systems and centralized control strategies, exhibit significant shortcomings in terms of response speed, control precision, and adaptability to complex operating scenarios11. Owing to their independently actuated wheel drives, DDEVs can substantially improve longitudinal stability through the real-time regulation of individual wheel torque. Nevertheless, the nonlinear and strongly coupled nature of vehicle dynamics, coupled with rapidly changing driving environments, renders traditional control approaches insufficient to meet advanced control demands12. Therefore, the development of efficient, robust, and adaptive longitudinal stability control strategies is not only critical for enhancing the dynamic performance of DDEVs, but also ensures the safety and comfort of intelligent vehicles under diverse operating conditions13,14. Longitudinal stability has been developed for distributed-drive vehicles (DDEVs) in recent years15,16. In light of increasing requirements for active safety, the development of more predictive control strategies has become essential17. Accordingly, control methods have been developed to extend response time by initiating braking interventions in advance, thereby enabling the implementation of more precise control algorithms18,19. The conventional electronic differential control strategy based on the Ackermann steering principle and logical thresholding is relatively simple, but its applicability is limited to low-speed and specific operating conditions. Furthermore, the necessity for empirically determined slip ratio thresholds constrains its real-time dynamic responsiveness20. By contrast, fuzzy control strategies have attracted increasing research interest owing to their rapid response capabilities21. For instance, fuzzy PID-based control methods can achieve stable control across a wide range of speeds without requiring precise system modeling22. Under wet or slippery road conditions, fuzzy logic combined with road adhesion coefficient estimation significantly enhances vehicle stability23. In addition, dual fuzzy control approaches, which implement differentiated anti-slip controllers for the front and rear axles, further address the challenges of tire adhesion performance affected by axle load transfer24. Integrating SMC with fuzzy adaptive PID strategies has proven effective in reducing braking distances in ABS control systems25, while SMC-based torque distribution methods can maintain slip ratios within 10%, thereby minimizing power loss induced by wheel slip26. Nevertheless, the inherent chattering problem associated with conventional SMC has stimulated substantial research into improvement methods. Notably, radial basis function (RBF) neural network-based SMC controllers can mitigate chattering through adaptive adjustment while preserving system robustness27. Additionally, convolutional neural networks (CNNs) have been implemented for road surface identification, enabling fuzzy control-based torque distribution, which effectively reduces energy consumption and extends driving range28. MPC demonstrates remarkable advantages in longitudinal stability control. MPC optimizes control performance under multiple constrained scenarios, particularly under high-speed conditions, where dynamic constraints play a critical role in stability control29. For example, tire slip ratio-based MPC controllers optimize stability across various operating conditions through real-time adjustment of objective function weights30. The multi-model MPC (MMPC) approach further enhances control robustness and smoothness under complex operating conditions by establishing multiple sub-models with dynamic switching capabilities31. Moreover, nonlinear MPC combined with optimized torque allocation ensures that all wheel slip ratios remain within ideal ranges, thereby improving overall vehicle stability32. Recent studies have also attempted to improve the computational efficiency and formulation of predictive control. For example, Wu et al. proposed a fast predictive functional control method based on auxiliary variable optimization and adopted an extended state-space model to avoid the use of additional observers when the required system variables can be directly measured and used as states33. This approach is advantageous because it can reduce observer-induced estimation errors and simplify the controller implementation when reliable measured state variables are available. However, in DDEV stability control, several key states required for control, such as lateral velocity, sideslip-related variables, road adhesion information, and reliable slip-related motion states, are difficult to measure directly with sufficient accuracy under complex driving conditions. Therefore, state estimation remains necessary in the present study to provide robust inputs for the controller. Nevertheless, the practical application of MPC in DDEV control continues to present significant challenges. On one hand, meeting real-time requirements necessitates the simplification of nonlinear vehicle dynamics models to reduce computational complexity, which may, however, compromise model accuracy. On the other hand, the generalization capability of MPC across diverse operating conditions remains limited, restricting its direct applicability in real-world engineering scenarios. Furthermore, achieving robust stability control at high speeds remains a research challenge, calling for further algorithmic optimization to ensure rapid response and high robustness. Despite these advances, significant gaps remain. Conventional MPC relies on single linearized models, limiting its generalization across varying driving scenarios. Robust MPC variants improve disturbance rejection but sacrifice optimality, while adaptive MPC requires extensive online parameter identification, increasing computational burden. Nonlinear MPC offers high accuracy but faces real-time implementation challenges due to computational complexity. In summary, although these have yielded notable advances in DDEV longitudinal stability control, further research is required to address the demands of complex driving conditions and real-time performance, particularly with respect to enhancing the generalizability of control algorithms and their adaptability to practical engineering applications. To address these challenges, this study focuses on key technologies for longitudinal stability control in DDEVs. First, a comprehensive vehicle dynamics model for DDEVs is established, including the body dynamics, drive motor, and tire force models; the accuracy is validated through simulation experiments. Second, a vehicle state estimation method is developed by integrating LiDAR and inertial measurement unit (IMU) data, where IMU is used for distortion correction of LiDAR data, and short-term temporal prediction of longitudinal and lateral acceleration is incorporated. A multi-model MPC-based longitudinal stability controller is then designed. Operating conditions are classified based on vehicle state parameters, and dedicated MPC controllers are developed for each scenario. The control outputs of the individual MPCs are fused using weighted coefficients that reflect their relevance to the current condition, ensuring both optimal longitudinal stability and smooth transitions between operating modes. Finally, a HIL experimental platform based on a steer-by-wire chassis is constructed to conduct integrated experiments on vehicle dynamics modeling, state estimation, and longitudinal stability control methods for DDEVs. System feasibility is validated by comparing the yaw rates, thereby providing a solid foundation for practical engineering applications. Compared with the extended state-space predictive control method reported by Wu et al.33, the proposed framework is designed for a different control scenario. The method in33 is suitable when the measured variables can be directly selected as system states, thereby avoiding observer-induced estimation errors. In contrast, the present DDEV stability-control problem involves unmeasured or indirectly measured states, sensor noise, tire slip, and road-adhesion variation. Therefore, the LiDAR-IMU-AUKF-TSA estimator is introduced to reconstruct reliable vehicle states before multi-model MPC optimization. The advantage of this observer-based structure is its applicability to complex vehicle dynamics and perception-limited conditions, whereas its disadvantage is that estimation errors may propagate to the controller. To reduce this effect, adaptive noise tuning and multi-sensor fusion are adopted in this study. Compared with existing switching MPC methods, the proposed controller does not select only one active model at each time instant. Instead, it fuses the control outputs of multiple scenario-specific MPC controllers according to their matching degrees with the current vehicle state, thereby reducing discontinuities during mode transitions. Compared with gain-scheduled MPC, the proposed method explicitly constructs multiple models corresponding to typical DDEV operating scenarios, rather than merely adjusting controller parameters according to scheduling variables. Therefore, the proposed framework provides a more scenario-aware and smoother control strategy for DDEV longitudinal stability control. Dynamics modeling of distributed drive electric vehicle Vehicle dynamics modelling The resulting DDEV dynamics model is illustrated in Fig. 1. On a flat road surface, the vehicle’s center of gravity is defined as the origin O of the vehicle coordinate system. The x, y, and z axes of the vehicle coordinate system are defined according to the right-hand rule. Vehicle dynamics model of the distributed drive electric vehicle. Accordingly, the longitudinal dynamics equation: The lateral kinematic equation of the vehicle is The yaw motion equation of the vehicle is Where the symbols are defined as follows: denote longitudinal tire forces of front-left, front-right, rear-left, and rear-right wheels, respectively; represent corresponding lateral tire forces; is vehicle mass; and are longitudinal and lateral accelerations; and are longitudinal and lateral velocities; is yaw rate; is yaw moment of inertia; a and b are distances from center of mass to front and rear axles; and are front and rear track widths; is steering angle; is sideslip angle. Tire modelling The tire serves as the sole medium for force transmission between the vehicle and the road, and its dynamic have a direct impact on vehicle stability control. The tire model implemented in CarSim exhibits characteristics consistent with the Magic Formula model. Therefore, the Magic Formula tire model is used in this study. The longitudinal dynamics equation of the vehicle is: Taking leftward steering as an example, the tire slip angles for the four wheels of the vehicle are: The longitudinal velocities at the centers of the four wheels are: To reduce computational complexity, it is assumed that under ideal conditions where wheel vertical motion and tire deformation are neglected, the wheel dynamics model is in Fig. 2. Wheel dynamic model. Therefore, the wheel dynamics equations and the wheel slip ratio are: Where ; represents the wheel’s moment of inertia; denotes the angular acceleration of wheel rotation; indicates the motor torque and r is the wheel radius. The longitudinal force and lateral force of the four wheels can be obtained. The expression of the Magic Formula tire model is: Where B represents the stiffness factor; C denotes the shape factor of the curve; D is the peak factor of the curve; E indicates the curvature factor of the curve.The variable x can represent either the wheel slip ratio or the tire slip angle . Correspondingly, y describes either the longitudinal force or the lateral force , depending on the coefficients and the choice of x. Specifically: When y represents the longitudinal force , x is the wheel slip ratio . When y represents the lateral force , x is the tire slip angle . The longitudinal forces for the four wheels are calculated as: The lateral forces for the four wheels are calculated as: Vehicle state estimation fusing LiDAR and IMU Conventional vehicle stability control systems usually estimate vehicle states using wheel-speed sensors and IMU. These sensors are effective for standard driving conditions, but their reliability may decrease under severe wheel slip or rapidly varying road adhesion. Wheel-speed-based estimation directly depends on tire rotation and can therefore be biased when the longitudinal slip ratio becomes large. Although IMU can provide high-frequency acceleration and angular-rate information, pure inertial estimation may suffer from drift and accumulated errors. To improve the robustness of state estimation under complex operating conditions, LiDAR is introduced as an exteroceptive sensor to provide relative pose and displacement information through point-cloud feature matching. Since this information is obtained from environmental geometric features rather than wheel rotation, it can serve as a complementary observation when wheel-speed-based estimation becomes unreliable. Therefore, LiDAR is used in this study not to replace conventional onboard sensors, but to enhance vehicle-state estimation robustness by fusing environmental geometric constraints with high-frequency IMU measurements. The vehicle state estimation system consists of two modules: a laser-inertial odometry module and a vehicle state estimation module. In the odometry module, to address point cloud distortion caused by vehicle motion, the LiDAR pose change is predicted using IMU measurements to mitigate motion-induced distortion effects on odometry accuracy34. Subsequently, feature extraction is performed on the point cloud, and inter-frame feature point matching enables pose estimation, providing vehicle localization information. In the vehicle state estimation module, to reduce model errors and suppress filter divergence, a noise adaptation mechanism is incorporated into the Unscented Kalman Filter (UKF). Additionally, TSA is employed to predict longitudinal and lateral accelerations, further minimizing estimation errors. Finally, simulation experiments are conducted using CARLA to validate. The framework of state estimation is in Fig. 3. Vehicle state estimation system. Design of LiDAR-inertial odometry Kinematics-based methods circumvent GPS limitations in confined spaces and low-frequency scenarios35,36. Laser SLAM systems, unrestricted by environment and providing high-frequency outputs, are better suited for full-operation stability control. This study proposes fusing LiDAR and IMU data to estimate longitudinal velocity , lateral velocity and yaw rate . In the proposed LiDAR-inertial odometry module, LiDAR provides frame-to-frame geometric constraints by extracting and matching point-cloud features, while IMU provides high-frequency angular velocity and acceleration information for motion distortion compensation. This complementary fusion is particularly important for stability control because the estimated longitudinal velocity, lateral velocity, and yaw rate should remain reliable even when wheel slip occurs. Compared with wheel-speed-only estimation, the LiDAR-derived relative motion observation is less sensitive to tire slip; compared with IMU-only estimation, LiDAR constrains inertial drift through environmental feature matching. The LiDAR-inertial odometry system employs a mechanical rotating LiDAR that outputs point cloud frames per scanning cycle. Vehicle motion causes intra-frame sensor pose variations, leading to motion distortion through coordinate origin drift within individual frames. Enhancing motion distortion correction accuracy requires intra-frame pose data at frequencies exceeding the LiDAR sampling rate. The IMU (sampling at ≥ 100 Hz) captures intra-frame motion with sufficient fidelity to enable high-precision local pose estimation for motion compensation. To address frame-to-frame matching errors induced by transient object absence, point cloud clustering is performed using an angle-based region growing algorithm. For consecutive points and in a scan ring, the angle between their vectors from the sensor origin is computed. If (empirically determined), the points belong to the same cluster. Clusters with fewer than points are discarded as noise. For 3D-to-2D projection, point clouds are projected onto the x-y plane using orthographic projection: ()→(), preserving horizontal geometric relationships for computational efficiency. Vehicle motion state estimation The pose transformation matrix characterizes the relative pose variation of the vehicle coordinate system between consecutive timesteps (k,k+1), representing the translational displacement and yaw angle change over the time interval Δt, Under the assumption of piecewise constant velocity motion between frames, the derivative of constructs a coarse-grained observation . Since LiDAR-inertial odometry module lacks refinement of , contains substantial noise and is unsuitable for vehicle longitudinal stability control. To enhance estimation accuracy, is refined using an UKF with adaptive noise tuning, while time-series analysis is integrated with IMU data for further optimization. The vehicle motion state is described by a discrete-time state-space model: where and denote the state vectors at time steps k and k-1, respectively; F is the state transition matrix; is the process noise with covariance Q; is the observation vector; H is the observation matrix; and is the measurement noise with covariance R. Although the above representation is linear in form, the AUKF framework is adopted to improve robustness against model uncertainty, sensor noise variation, and nonlinear vehicle-motion characteristics in practical driving conditions. The UKF with adaptive noise tuning is subsequently applied through phase-based processing, commencing with state initialization and progressing to noise-adapted state estimation. Initialization The initial values of the system state vector and the error covariance are defined as follows: Subsequently, Sigma sampling points for the state vector are generated: Where n is the dimension of X, is the scaling parameter (typically ), and . Prediction update The Sigma points are propagated through the state transition model F: The predicted state and the error covariance at time k are then computed via weighted sums: where the weights are given by: Measurement update Sigma points for the measurement are calculated as follows: Where . The Sigma points are then mapped through the observation model H: The predicted measurement, the measurement error covariance, and the cross-covariance are given by: The Kalman gain is then computed as: The updated state and covariance at time k are: Noise covariance update Imprecisions in the state equation contribute to system noise, thereby affecting estimation accuracy and potentially causing filter divergence. The adaptive noise covariance adjustment operates as follows: The sliding window length is set to M=10 to balance responsiveness and noise suppression. The adjustment factor in Eq. (40) ensures decreases only when , preventing filter divergence from over-adjustment. The trace ratio quantifies the discrepancy between predicted and actual innovation covariances. This mechanism enables real-time compensation for model uncertainties and unmodeled dynamics.To mitigate this, the difference between the observed and predicted measurement is defined as: Given that d is influenced by both model errors and initial estimation inaccuracies, the covariance of d can be used to adaptively estimate the system noise covariance. The theoretical covariance of d is: Due to measurement noise, the empirical covariance of d may deviate from the theoretical value. The empirical covariance is thus calculated as: Where M denotes the length of the sliding window. Comparing the theoretical and empirical covariances enables adaptive adjustment of the system noise covariance , if , is decreased; if , is not adjusted to prevent filter divergence. The adjustment factor is defined as: The noise covariance matrices are updated: Longitudinal and lateral acceleration prediction In accordance with Eq. (21), prediction of both longitudinal and lateral vehicle accelerations is required. As these accelerations are subject to significant variability due to driver behavior and road conditions, accurate mathematical modeling using deterministic methods is challenging. For longitudinal and lateral acceleration prediction, ARIMAmodels are employed with parameters , determined via Akaike Information Criterion (AIC) optimization on training data. The model uses historical samples (corresponding to 0.1 s at 100 Hz IMU sampling) to predict the acceleration at the next timestep. The autoregressive coefficients and moving average coefficients for longitudinal acceleration;, for lateral acceleration, identified from CARLA simulation data. The ARIMA model enables short-term prediction by incorporating p consecutive historical observations andq consecutive historical random disturbances. Its mathematical formulation is given by: Where p and q represent the orders of the model; denotes the predicted value at time, are the autoregressive coefficients; are sequential historical noise terms; and are the moving average coefficients. In this study, sets of ten consecutive longitudinal or lateral acceleration values are used as historical input data for the ARIMA model to predict the acceleration at the subsequent time step. Figure 4 compares the predicted accelerations with the actual measured values, demonstrating that the ARIMA model provides reliable predictive performance for both longitudinal and lateral accelerations. Comparison of acceleration prediction results. (a) Prediction results of longitudinal acceleration. (b) Prediction results of lateral acceleration. Simulation analysis of vehicle state estimation Simulation experiments were conducted using the CARLA platform. Vehicle speeds were set within the range of 0–30 m/s. The vehicle driving trajectory is shows in Fig. 5(a). Three approaches were comparatively analysed: (1) results based solely on differentiated LiDAR-inertial odometry, (2) results obtained by combining LiDAR-inertial odometry with AUKF, and (3) results integrating time series analysis into the AUKF framework (TSA-AUKF). Comparisons of the estimated values with reference values provided by CARLA are presented in Fig. 5. Figure 5(b) is longitudinal acceleration prediction comparison. Figure 5(c) is lateral acceleration prediction comparison. Figure 5(d) is yaw rate prediction comparison, the TSA-AUKF is closest to the true value. Comparison of vehicle state estimation results: (a) vehicle trajectory; (b) longitudinal velocity estimation; (c) lateral velocity estimation; and (d) yaw-rate estimation. To provide a more intuitive assessment of performance differences, error analysis was conducted. The error curves, relative to the CARLA reference values, are shown in Fig. 6. Comparative Error Profiles of Motion States. (a) Longitudinal Velocity Error. (b) Lateral Velocity Error. (c) Yaw Rate Error. RMSE was used to quantitatively assess the estimation accuracy of each method, as summarized in Table 1. For longitudinal velocity estimation, the RMSEs for LiDAR-inertial odometry, AUKF, and TSA-AUKF were 2.444 m/s, 1.569 m/s, and 1.371 m/s, respectively. For lateral velocity, the corresponding RMSEs were 0.360 m/s, 0.333 m/s, and 0.243 m/s. For yaw rate, the RMSEs were 0.407 deg/s, 0.245 deg/s, and 0.111 deg/s, respectively. TSA-AUKF achieves the lowest estimation error. Table 1Error Results.Error MetricLiDAR-Inertial OdometryAUKFTSA-AUKFLongitudinal Velocity (m/s)2.4441.5691.371Lateral Velocity (m/s)0.3600.3330.243Yaw Rate (deg/s)0.4070.2450.111 Error Results. Compared with LiDAR-inertial odometry, TSA-AUKF improved the estimation accuracy by approximately 44%, 33%, and 73%, respectively. When compared to AUKF, the improvements achieved by TSA-AUKF were approximately 13%, 27%, and 55%, respectively. These results indicate that the LiDAR-inertial odometry provides useful external motion observations, while the AUKF and TSA modules further refine the raw odometry output and reduce estimation errors. Therefore, the advantage of introducing LiDAR is not only reflected in localization, but also in providing slip-independent relative motion information for vehicle-state estimation, which improves the reliability of the subsequent longitudinal stability controller. Multi-model MPC-based longitudinal stability control method The proposed multi-model MPC is designed to address the limitations of single-model MPC, switching MPC, and gain-scheduled MPC under complex DDEV operating conditions. Single-model MPC may suffer from reduced accuracy when the vehicle operates far from the nominal linearization point. Switching MPC can improve adaptability by selecting different models, but abrupt switching may cause discontinuous control actions. Gain-scheduled MPC adjusts controller parameters according to scheduling variables, but it may not fully capture the nonlinear differences among typical vehicle operating modes. To overcome these limitations, this study establishes multiple scenario-specific MPC controllers and integrates their outputs through a matching-degree-based weighted fusion strategy. Based on the previously established high-frequency and high-precision estimation of vehicle states and road adhesion coefficients, this chapter develops a multi-model MPC approach for longitudinal stability control of DDEV under varying driving conditions37. Multiple MPC controllers are designed for distinct scenarios to generate optimal stability control inputs. A neural network is introduced to identify the current driving condition, and a weighted fusion strategy integrates the optimal control outputs from each model according to the scenario matching degree. The proposed method is validated through CarSim/CARLA simulations. Multi-model MPC controller design Figure 7 presents the structure of the multi-model MPC controller. The state vector x, comprising and the optimal slip ratios for all four wheels, as input to the controller. Each scenario specific MPC computes the corresponding optimal control actions in real time. Flowchart of the multi-model MPC controller. MPC predicts the system dynamics over a horizon of N steps based on the current state, solves an online constrained optimization problem, and applies control input at each time step. This receding horizon strategy enables MPC to generate optimal control actions under various constraints and system requirements. To improve reproducibility, the MPC formulation used in each scenario-specific controller is defined as follows. For the i-th operating scenario, the discrete prediction model is expressed as where denotes the scenario index. The state vector is defined as where , , and rdenote longitudinal velocity, lateral velocity, and yaw rate, respectively; , , , and denote the slip ratios of the front-left, front-right, rear-left, and rear-right wheels. The inclusion of and rin the state vector does not change the main objective of longitudinal stability control. Instead, these variables are introduced to monitor the coupled lateral response caused by longitudinal torque allocation. In DDEVs, independent wheel torques directly affect the longitudinal tire forces at each wheel. If the left-right longitudinal forces are unbalanced, an additional yaw moment may be generated, leading to yaw-rate deviation and sideslip. Therefore, and rare included as auxiliary stability states to ensure that longitudinal torque regulation does not deteriorate overall vehicle stability. The control input is defined as where , , , and are the driving/braking torques of the four in-wheel motors. The measured disturbance vector includes the steering angle and the estimated road adhesion coefficient. The optimization problem of the i-th MPC controller is formulated as Here, and are the state-error and control-increment weighting matrices, respectively. and denote the prediction horizon and control horizon. In this study, the values of , , , , and the torque/slip-ratio limits are summarized in Table 2. Table 2Error comparison results.ParameterConventional MPCMulti-model MPCSideslip angle of the center of mass0.136 deg0.032 degYaw rate0.314 deg/s0.096 deg/sSlip ratio of the front left wheel0.0150.009Slip ratio of the rear left wheel0.0080.005Slip ratio of the front right wheel0.0140.009Slip ratio of the rear right wheel0.0090.005 Error comparison results. Table 2Main parameters of the multi-model MPC controller.ParameterMeaningValue used in this study Prediction horizon20 Control horizon5 State-error weighting matrixdiag(10, 20, 80, 150, 150, 150, 150) Control-increment weighting matrixdiag(0.01, 0.01, 0.01, 0.01) Slip-ratio limits[− 0.10, 0.10] Motor torque limits[− 1200, 1200] N⋅mΔTorque-rate limits[− 200, 200] N⋅m Main parameters of the multi-model MPC controller. Multi-scenario vehicle state dataset establishment The quality of the dataset is vital for neural networks, as it directly impacts model performance. Typically, data are transformed into matrix or vector forms to facilitate computational processing and to mitigate issues such as gradient explosion and vanishing gradients. The dataset used for training the scenario-identification neural network was generated from vehicle dynamics simulations under representative DDEV operating conditions. The simulation scenarios covered acceleration and deceleration maneuvers, low- and high-speed ranges, and small- and large-sideslip conditions. These conditions correspond to the eight operating scenarios used in the multi-model MPC. Each sample was constructed from the vehicle state vector used for scenario identification, including longitudinal velocity, lateral velocity, yaw rate, wheel slip ratios, vehicle speed, and sideslip angle used for threshold-based scenario classification. The corresponding label of each sample was determined according to the operating-condition definitions described below and summarized in Table 3. A total of 120,000 samples were collected from 15 simulation runs, each lasting 80 s with a sampling time of 0.01 s. The dataset was randomly divided into training, validation, and test sets with a ratio of 70%, 15%, and 15%, corresponding to 84,000, 18,000, and 18,000 samples, respectively. The training set was used to update the network weights, the validation set was used for hyperparameter tuning and overfitting monitoring, and the test set was used only for final performance evaluation. Before training, all input variables were normalized to the range [0,1] to reduce the influence of different physical units and magnitudes. Table 3Operating conditions and corresponding prediction models in the multi-model MPC.Model indexLongitudinal conditionSpeed conditionSideslip conditionTire-force approximation M 1 DecelerationLow speedSmall sideslipLinear tire-force model M 2 DecelerationLow speedLarge sideslipCubic nonlinear tire-force model M 3 DecelerationHigh speedSmall sideslipLinear tire-force model M 4 DecelerationHigh speedLarge sideslipCubic nonlinear tire-force model M 5 AccelerationLow speedSmall sideslipLinear tire-force model M 6 AccelerationLow speedLarge sideslipCubic nonlinear tire-force model M 7 AccelerationHigh speedSmall sideslipLinear tire-force model M 8 AccelerationHigh speedLarge sideslipCubic nonlinear tire-force model Operating conditions and corresponding prediction models in the multi-model MPC. The longitudinal operational states of the DDEV are classified into three representative scenarios: Acceleration: for all (wheel slip rate derivatives decrease, indicating traction increase); Acceleration: for all (wheel slip rate derivatives decrease, indicating traction increase); Constant speed: for all wheels (slip rates near optimal); Constant speed: for all wheels (slip rates near optimal); Deceleration: for all wheels (slip rate derivatives increase, indicating braking). Deceleration: for all wheels (slip rate derivatives increase, indicating braking). The speed threshold distinguishes low/high speed regimes based on the nonlinear tire force transition region. The sideslip angle threshold marks the linear/nonlinear tire force boundary per the Magic Formula model. The speed threshold km/h was selected according to the simulation speed range and the change in vehicle dynamic response. In this study, the vehicle speed range for scenario generation is 0–90 km/h; therefore, 45 km/h is used as the midpoint to separate low-speed and high-speed regimes. More importantly, this boundary reflects the fact that the vehicle yaw and lateral responses become more sensitive to steering input and tire-force variation at higher speeds. Thus, the high-speed scenarios require different prediction models and control weighting from the low-speed scenarios. The sideslip-angle threshold rad was selected according to the tire-force characteristics. When the wheel sideslip angle satisfies rad, , the lateral tire force remains close to the linear region of the Magic Formula tire model. Therefore, the linear tire-force approximation is used for small-sideslip scenarios. When rad, the tire-force nonlinearity becomes more evident, and the cubic polynomial approximation is used for large-sideslip scenarios. The upper bound of 0.02 rad was selected to cover large-sideslip but still controllable operating conditions in the designed simulation maneuvers. These thresholds are model-dependent and can be recalibrated for different vehicles, tires, or road-adhesion conditions. These scenarios are defined to evaluate the longitudinal dynamics of the DDEV. The state equations for acceleration and constant-speed modes are identical; therefore, these two conditions are merged. Consequently, the longitudinal operational modes are reduced to acceleration and deceleration. For lateral dynamics, four representative scenarios are defined according to the velocity and wheel sideslip angles: It should be noted that the above boundaries are not universal constants. They are selected for the vehicle model and maneuver range considered in this study, based on the speed distribution, the linear/nonlinear transition of the tire-force characteristics, and the requirement that the large-sideslip cases remain within a controllable stability region. For another vehicle platform or tire-road condition, these thresholds can be recalibrated using the corresponding tire model and experimental data. The lateral tire force adopts different expressions under these operating modes. By performing a Taylor expansion, a cubic polynomial (Eq. 48) is obtained. As large, the system enters a nonlinear region, and Eq. 44 approximates the Magic Formula well. As small, the system behaves linearly, and Eq. 49 provides a close approximation. This allows for simplification of the Magic Formula as follows: Thus, Eq. 48 is used for large sideslip angles, and Eq. 49 for small sideslip angles. Combining the two longitudinal and four lateral operational modes yields eight typical driving scenarios: deceleration low speed small sideslip, deceleration low speed large sideslip, deceleration high speed small sideslip, deceleration high speed large sideslip, acceleration low speed small sideslip, acceleration low speed large sideslip, acceleration high speed small sideslip, and acceleration high speed large sideslip. To make the model set clearer, the eight prediction models used in the multi-model MPC are summarized in Table 3. In the Table 3, the low-/high-speed boundary and the small-/large-sideslip boundary are determined by the velocity threshold and sideslip-angle threshold defined in the section “Multi-scenario vehicle state dataset establishment”. For small sideslip angles, the linear approximation of the Magic Formula tire model is used, whereas for large sideslip angles, the cubic polynomial approximation is adopted. Weighted fusion of multi-model control inputs Based on the previous sections, the control inputs from the multi-model MPC under eight typical operating scenarios, as well as the corresponding scenario matching weights, have been obtained. These results require further processing to determine the final control command. In real-world driving, the future operating condition is inherently uncertain; however, since the vehicle state equation has been formulated in differential form, the next state is influenced by the current state. Therefore, it is inappropriate to simply discard control inputs associated with lower matching weights. The neural network for scenario identification comprises three fully-connected layers: input layer (8 neurons for state vector ), hidden layer (32 neurons with ReLU activation), and output layer (8 neurons with softmax activation representing scenario matching probabilities ). The weighted fusion computes final control input , where . This ensures smooth control transitions when vary gradually, avoiding abrupt switching. This weighted-fusion mechanism is a key difference between the proposed method and conventional switching MPC. In switching MPC, the controller output is usually generated by a single selected model, which may lead to abrupt changes when the recognized operating mode changes. In contrast, the proposed method retains the contributions of all scenario-specific controllers according to their matching probabilities. As a result, the final control command varies continuously with the vehicle state, improving transition smoothness and control robustness. The neural network was trained using the cross-entropy loss function, because the scenario-identification task is formulated as an eight-class classification problem. The Adam optimizer was used for parameter updating, with an initial learning rate of 0.001. The maximum number of training epochs was set to 100, and the mini-batch size was set to 256. The network parameters corresponding to the lowest validation loss were selected as the final model to avoid overfitting. After training, the softmax output was used as the scenario matching-degree vector for the weighted-fusion multi-model MPC. The matching-degree vector generated by the softmax layer is expressed as where denotes the matching degree between the current vehicle state and the i-th operating scenario. The weights satisfy Each scenario-specific MPC controller calculates an optimal control input . The final control command applied to the vehicle is obtained by weighted fusion: Therefore, all scenario-specific MPC controllers contribute to the final control command according to their matching degrees, rather than selecting only one active controller. The input constraints are preserved after weighted fusion because the fusion operation is a convex combination of feasible control inputs. Let the admissible input set be defined as Each scenario-specific MPC solves its optimization problem subject to . Since is a convex set and the matching weights satisfy and , the fused input also satisfies . Therefore, the motor torque and torque-increment constraints remain satisfied after weighted fusion. For the stability requirements, the slip-ratio bounds and stability-related states are considered in the scenario-specific MPC formulation. After weighted fusion, a post-fusion safety check is further performed. The fused torque command is first checked against the admissible torque and torque-increment bounds. Then, a one-step prediction is used to examine whether the wheel slip ratios remain within the stable range. If a violation is detected, the fused command is projected back into the admissible set: and is applied to the four in-wheel motors. If no violation is detected, is directly applied. This mechanism ensures that the weighted-fusion strategy maintains input feasibility while preserving the slip-ratio stability requirement. Accordingly, a direct weighted fusion approach is employed, whereby the control inputs from all eight operating scenarios are aggregated based on their respective matching weights. This strategy maximizes the utilization of available information and ensures smooth transitions during mode switching. The weighted fusion process is illustrated in Fig. 8. The weighted fusion scheme. Online implementation of the proposed LiDAR-IMU-AUKF-TSA-based weighted-fusion multi-model MPC strategy. Simulation analysis The simulation experiments were conducted under a serpentine maneuver scenario. The initial velocity was set to , with a steering angle . The road adhesion coefficient was maintained at 0.8 for 0–40 s and reduced to 0.5 for 40–80 s. The longitudinal velocity decreased uniformly from 80 km/h to 10 km/h over 0–20 s, increased back to 80 km/h from 20 to 40 s, decreased again to 10 km/h from 40 to 60 s, and finally increased to 80 km/h from 60 to 80 s. The total simulation time was 80 s. To clarify the experimental comparison, the control strategies considered in this study are further explained. The conventional MPC uses a fixed prediction model and fixed controller parameters. The single-model MPC uses one nominal DDEV prediction model throughout the entire maneuver. The switching MPC selects only one scenario-specific MPC according to the maximum scenario matching degree. In contrast, the proposed multi-model MPC fuses the optimal control inputs of all scenario-specific MPC controllers according to their matching degrees. Therefore, the proposed method differs from switching MPC by avoiding hard switching between controllers and improving the smoothness of control transitions under varying operating conditions. Figure 9 illustrates the scenario matching weights output by the neural network. It can be observed that the matching weights dynamically adjust according to the changing driving conditions. Variation of matching degree weighting coefficients over time. Figure 10 compares the vehicle states between the multi-model MPC and conventional MPC. As shown in Fig. 10(a), the multi-model MPC demonstrates significantly improved stability compared to the conventional MPC, effectively maintaining vehicle stability. The differences in yaw rate, depicted in Fig. 10(b), are less pronounced and will be further analyzed in the subsequent error analysis. Figures 10(c-f) indicate that the slip ratios of all four wheels remain within stable regions. In this study, the slip ratios of the four wheels are regarded as the primary indicators of longitudinal stability because they directly represent longitudinal tire-force utilization. The yaw rate and sideslip angle are used as coupled stability indicators to evaluate whether the longitudinal torque distribution causes undesired yaw or lateral motion. Therefore, the combined evaluation of slip ratio, yaw rate, and sideslip angle can reflect both the longitudinal tire stability and the coupled vehicle attitude stability of the DDEV. Comparison of simulation results. (a) Comparison of centroid sideslip angle (b) Comparison of yaw rate. (c) Comparison of left front wheel slip ratio. (d) Comparison of left rear wheel slip ratio. (e) Comparison of right front wheel slip ratio. (f) Comparison of right rear wheel slip ratio. Figure 11 shows the errors of the multi-model MPC and conventional MPC compared to the CarSim reference values. Error curves of simulation results. (a) Centroid sideslip angle error. (b) Yaw rate error. (c) Left front wheel slip ratio error. (d) Left rear wheel slip ratio error. (e) Right front wheel slip ratio error. (f) Right rear wheel slip ratio error.. RMSE analysis for the sideslip angle and yaw rate is summarized in Table 2. Under conventional MPC control, the RMSEs for the sideslip angle and yaw rate are 0.136 deg and 0.314 deg/s, respectively, while the RMSEs for the slip ratios of the four wheels are 0.015, 0.008, 0.014, and 0.009. With the multi-model MPC, the RMSEs for the sideslip angle and yaw rate are reduced to 0.032 deg and 0.096 deg/s, and the slip ratios are maintained at 0.009, 0.005, 0.009, and 0.005, respectively. These comparative results demonstrate that the proposed multi-model MPC method significantly enhances vehicle longitudinal stability. Experimental validation of the multi-model MPC-based longitudinal stability control system Experimental platform setup The HIL platform consists of a real-time vehicle dynamics simulation unit, a multi-model MPC controller unit, and a communication interface. The vehicle dynamics and road environment are simulated in real time, while the proposed controller is deployed on the controller unit to calculate four-wheel driving/braking torque commands. The estimated vehicle states, including longitudinal velocity, lateral velocity, yaw rate, and wheel slip ratios, are transmitted to the controller at each sampling instant. The controller then sends the optimized torque commands of the four in-wheel motors back to the vehicle dynamics simulation unit. The sampling time of the control system is set to 0.01 s, consistent with the MPC implementation. The experimental vehicle is equipped with a drive-by-wire chassis featuring four in-wheel motors, consistent with the DDEV configuration and suitable for the study’s requirements. The tires used are model 215/45ZR17. All computational and sensing equipment is mounted on the chassis, as illustrated in Fig. 12. The sensor suite includes a 16-line mechanical LiDAR, an IMU, and a high-precision MEMS integrated navigation system, highlighted by the red box in the Fig. 12. The MEMS navigation system (CGI-430) offers centimeter-level positioning accuracy and a 10 Hz data acquisition rate, providing reference values for vehicle position, yaw rate, and longitudinal velocity. Experimental platform. The LiDAR SLAM framework is implemented using the Robot Operating System (ROS), which supports modular, point-to-point communication via TCP. Both asynchronous topic-based and synchronous service-based communication modes are utilized, with topic-based communication being predominant. The entire control and estimation software stack is ported to ROS, and all program modules are developed in accordance with ROS conventions. Experimental results and analysis Considering the constraints of the campus environment, the maximum speed of the experimental platform was limited to 15 km/h. To ensure consistency across comparative tests, all experiments were conducted in autonomous mode with a predetermined route based on a high-definition map and marked starting points, following a double lane-change trajectory. The actual vehicle path is shown in Fig. 13. Vehicle trajectory. The variation of scenario matching weights is depicted in Fig. 14. Time-Varying Matching Coefficient Weight Distribution under Multi-Condition Driving Scenarios. Figure 15(a) compares the reference longitudinal velocity with the value estimated by TSA-AUKF. The average error is only 1.612%, indicating high estimation accuracy. Figure 15(b) presents the yaw rate controlled by conventional MPC and multi-model MPC, compared with reference values. As the curves are visually similar, further quantitative error analysis is provided below. Figure 15(c) records the slip ratios of all four wheels under multi-model MPC control, all maintained within the stable range of [−0.1, 0.1]. Figure 15(d) shows the estimated road adhesion coefficient. As the experiment was conducted on dry cement pavement, the expected coefficient is within 0.7–0.9, and the estimated values are consistent with this range. Experimental Comparison Results of Control Strategies. Figure 16 presents the yaw rate errors relative to the reference for both conventional and multi-model MPC controllers. In the HIL experiment, conventional MPC was selected as the main baseline because it represents a widely used constrained optimal control approach with a single prediction model and fixed parameters. The comparison focuses on verifying whether the proposed controller can achieve real-time implementation and improve stability performance over a representative MPC baseline on the experimental platform. The additional discussion of single-model MPC and switching MPC in the simulation section is used to clarify the methodological difference between the proposed weighted-fusion strategy and other model-selection strategies. RMSE analysis shows that under conventional MPC, the yaw rate RMSE is 1.174 deg/s, while under multi-model MPC, the RMSE is reduced to 0.919 deg/s, representing a 21.7% improvement in performance. Although yaw rate is commonly used as a lateral dynamic metric, it is also relevant to the proposed longitudinal stability controller because the controller regulates the longitudinal driving/braking torques of four independent wheels. Any imbalance in the left-right longitudinal tire forces can generate an additional yaw moment. Therefore, the yaw-rate error in the HIL/vehicle experiment is used to evaluate whether the proposed longitudinal torque allocation strategy can suppress undesired yaw response while maintaining wheel slip within the stable range. Yaw rate error curve. Conclusion This paper addresses the challenges of inadequate vehicle state estimation accuracy and limited generalization capability of conventional MPC in complex driving scenarios. The main contributions and quantitative outcomes are summarized as follows: Development and Validation of a DDEV Seven-Degree-of-Freedom Model: A comprehensive seven-degree-of-freedom dynamic model tailored for DDEV was constructed, incorporating a three-degree-of-freedom vehicle body and four independent wheel models. The model’s accuracy and applicability were validated through co-simulation with CarSim and Simulink. Development and Validation of a DDEV Seven-Degree-of-Freedom Model: A comprehensive seven-degree-of-freedom dynamic model tailored for DDEV was constructed, incorporating a three-degree-of-freedom vehicle body and four independent wheel models. The model’s accuracy and applicability were validated through co-simulation with CarSim and Simulink. LiDAR-IMU Fusion-Based State Estimation Method: To overcome the limitations of tire-force-based and GPS-dependent estimation, a LiDAR-IMU fusion algorithm was proposed. The method integrates adaptive noise estimation and time series analysis, yielding high estimation accuracy. CARLA-based simulations demonstrated RMSEs of 1.371 m/s for longitudinal velocity, 0.243 m/s for lateral velocity, and 0.111 deg/s for yaw rate, indicating robust performance under diverse conditions. LiDAR-IMU Fusion-Based State Estimation Method: To overcome the limitations of tire-force-based and GPS-dependent estimation, a LiDAR-IMU fusion algorithm was proposed. The method integrates adaptive noise estimation and time series analysis, yielding high estimation accuracy. CARLA-based simulations demonstrated RMSEs of 1.371 m/s for longitudinal velocity, 0.243 m/s for lateral velocity, and 0.111 deg/s for yaw rate, indicating robust performance under diverse conditions. Multi-Model MPC Longitudinal Stability Controller: A multi-model MPC framework was developed, utilizing scenario classification via a neural network to match appropriate vehicle dynamics models in real time. This approach ensures smooth transitions across operational modes and enhances control robustness. Simulation results show that the proposed controller achieves an RMSE of 0.032 deg for sideslip angle and 0.096 deg/s for yaw rate, significantly outperforming conventional MPC. Multi-Model MPC Longitudinal Stability Controller: A multi-model MPC framework was developed, utilizing scenario classification via a neural network to match appropriate vehicle dynamics models in real time. This approach ensures smooth transitions across operational modes and enhances control robustness. Simulation results show that the proposed controller achieves an RMSE of 0.032 deg for sideslip angle and 0.096 deg/s for yaw rate, significantly outperforming conventional MPC. Experimental Validation on a Real Vehicle Platform: Real-vehicle tests were conducted on a drive-by-wire platform under double lane-change scenarios. The proposed state estimator achieved an average longitudinal velocity estimation error of just 1.612%. The multi-model MPC controller reduced the yaw rate RMSE to 0.919 deg/s, a 21.7% improvement over the conventional MPC. All wheel slip ratios remained within the stable range of [−0.1, 0.1], and the estimated road adhesion coefficients were consistent with actual surface conditions. Experimental Validation on a Real Vehicle Platform: Real-vehicle tests were conducted on a drive-by-wire platform under double lane-change scenarios. The proposed state estimator achieved an average longitudinal velocity estimation error of just 1.612%. The multi-model MPC controller reduced the yaw rate RMSE to 0.919 deg/s, a 21.7% improvement over the conventional MPC. All wheel slip ratios remained within the stable range of [−0.1, 0.1], and the estimated road adhesion coefficients were consistent with actual surface conditions. In addition, the proposed framework clarifies a methodological distinction from switching MPC and gain-scheduled MPC. By combining scenario-specific modeling with matching-degree-based weighted fusion, the controller improves adaptability to complex operating conditions while avoiding abrupt control switching. This provides a practical and reproducible control framework for longitudinal stability enhancement of DDEVs. The proposed system demonstrates significant improvements in longitudinal stability control; however, several limitations warrant consideration. Although the LiDAR-IMU-AUKF-TSA estimator improves the availability and robustness of vehicle states, estimation errors may still affect the MPC performance. Future work will investigate the integration of extended state-space predictive control and enhanced MPC formulations to reduce observer dependence and improve computational efficiency. First, the multi-model MPC requires pre-definition of scenario boundaries, which may not cover all extreme conditions. Second, while computational efficiency is maintained (average 4.8 ms per control cycle on Jetson AGX Xavier), further optimization is needed for resource-constrained embedded systems. Third, experimental validation was limited to low-speed scenarios due to campus safety constraints; high-speed performance requires additional verification. Future work will focus on: (1) extending the scenario library using data-driven clustering, (2) implementing hardware-in-the-loop testing for high-speed validation, and (3) developing fault-tolerant mechanisms for sensor failures. Acknowledgements This work was supported in part by the 2025 Youth Innovation Team Construction Scientific Research Program of the Shaanxi Provincial Education Department under Grant 25JP110, and in part by the Shaanxi Innovation Talent Promotion Plan-Science and Technology Innovation Team under Grant 2021TD-27. Author contributions Conceptualization, M.D. and C.Z.; methodology, M.D. and C.Z.; software, Z.G.; validation, M.D. and C.Z.; formal analysis, J.W.; investigation, C.Z.; resources, C.Z.; data curation, M.D.; writing—original draft preparation, J.W.; writing—review and editing, M.D.; visualization, M.D.; supervision, M.D.; project administration, M.D. Funding No Funding. Data availability Data sets generated during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data sets generated during the current study are available from the corresponding author on reasonable request.