Energy management method for pure electric tractor based on physical constraint neural network model

Abstract

This invention provides an energy management method for pure electric tractors based on a physically constrained neural network model, relating to the interdisciplinary field of intelligent control in agricultural machinery and new energy vehicles. The method involves training a physically constrained neural network model using a comprehensive loss function based on multi-dimensional spatiotemporal feature vectors, historical state sequences, and sensor-measured power data. The trained model is then used to calculate a predicted power demand sequence. A predictive control algorithm based on the physically constrained neural network model is then used for rolling optimization to obtain the optimal power allocation command for the pure electric tractor. Energy management of the pure electric tractor is then performed according to this optimal power allocation command. This invention solves the problem that existing purely data-driven predictive physically constrained neural network models, lacking physical mechanism constraints, lead to distorted power predictions and consequently, system energy allocation instability under variable load operating environments of pure electric tractors.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for agricultural machinery and new energy vehicles, and in particular to an energy management method for pure electric tractors based on a physical constraint neural network model. Background Technology

[0002] As modern agriculture transforms towards green and low-carbon practices, pure electric tractors, with their zero emissions, low noise, and high energy conversion efficiency, are gradually becoming an important development direction in the agricultural equipment field. However, unlike conventional electric vehicles that travel on paved roads, pure electric tractors face extremely complex working environments in farmland operations (such as deep tillage, plowing, and rotary tillage). Due to real-time nonlinear changes in soil resistivity, moisture content, surface undulation, and working depth, the tractor load exhibits strong randomness and severe transient fluctuations. This drastically fluctuating power demand poses a significant challenge to the energy management system (EMS) of pure electric tractors. Traditional energy management strategies (such as rule-based logic threshold strategies and power following strategies) are "reactive" controls, meaning they only adjust the battery output current after the sensor detects a change in load. This lag, especially during sudden load changes in farmland operations, easily leads to frequent high-rate current surges in the battery, severely shortening its cycle life and potentially causing localized thermal runaway. To achieve "predictive" energy management, existing technologies have begun to introduce time-series prediction algorithms (such as Long Short-Term Memory networks (LSTM) and Backpropagation (BP) neural networks) to predict the power demand of tractors in advance. However, these purely data-driven "black box" neural networks have a fatal flaw: they rely entirely on the statistical regularities of the training data and lack awareness of the physical kinematics and dynamic boundaries of the tractor. When strong vibrations in farmland cause noise in the sensor data, or when encountering extreme working conditions that have not been trained on (such as the plow teeth suddenly getting stuck on underground rocks), the prediction results given by the purely data-driven, physically constrained neural network models often violate basic physical laws (e.g., predicting extremely low power demand under rapid acceleration, or predicting torque exceeding the physical external characteristic limits of the motor). This will cause the downstream energy distribution controller to issue incorrect commands, leading to system shutdown or even hardware damage. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides an energy management method for pure electric tractors based on a physical constraint neural network model. This method solves the problem that existing pure data-driven predictive physical constraint neural network models, under variable load operating conditions of pure electric tractors, suffer from power prediction distortion due to the lack of physical mechanism constraints, which in turn leads to instability in system energy distribution.

[0004] To achieve the aforementioned objectives, the technical solution adopted by this invention is: an energy management method for pure electric tractors based on a physical constraint neural network model, comprising: S1: Based on the multi-source sensor data collected by the pure electric tractor, feature extraction is performed using the longitudinal dynamic equation and the machine coupling resistance equation to obtain a multi-dimensional spatiotemporal feature vector; S2: Using a multilayer perceptron as the baseline feedforward network, a deep physical constraint neural network model architecture is constructed to obtain the physical constraint neural network model; S3: Based on multidimensional spatiotemporal feature vectors, historical state sequences and sensor measured power data, combined with longitudinal dynamic equations, and using a comprehensive loss function, the physical constraint neural network model is trained through backpropagation and automatic differentiation techniques to obtain a trained physical constraint neural network model. S4: Based on the real-time acquired state sequence of the pure electric tractor, the trained physical constraint neural network model is used to calculate the future speed state sequence and the future demand power prediction sequence within the finite time domain. S5: Based on the future speed state sequence and demand power prediction sequence, the rolling optimization solution is performed using the physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command for the pure electric tractor. Energy management is then performed on the pure electric tractor according to the optimal power allocation command to obtain the energy management result and complete the energy management of the pure electric tractor.

[0005] Further, S1 includes: Based on multi-source sensor data collected by pure electric tractors, the longitudinal dynamic equation, which includes rolling resistance, slope resistance, acceleration inertial resistance, and implement operation resistance, is used to extract features from the implement coupling resistance equation, which includes static, first-order and second-order speed soil cutting drag coefficients. This yields a multi-dimensional spatiotemporal feature vector containing travel speed, slip ratio, slope angle, and implement traction resistance.

[0006] Further, S3 includes: Based on multidimensional spatiotemporal feature vectors, historical state sequences and sensor measured power data, the core resistance component of total driving resistance data is extracted by combining longitudinal dynamic equations, and a physical residual operator is constructed. The acceleration term is calculated by directly differentiating the future velocity state sequence output by the physical constraint neural network model with respect to continuous time variables using an automatic differentiation mechanism to construct a computational graph. The physical law residual terms are obtained by using the physical residual operator and the acceleration term; By using a comprehensive loss function that includes data fitting residuals, physical law residuals, spatiotemporal dynamic equilibrium constraint residuals, and operational boundary constraint residuals, a physical constraint neural network model is trained using backpropagation and automatic differentiation techniques, resulting in a well-trained physical constraint neural network model.

[0007] Furthermore, in the longitudinal dynamics equation, the expression for the total driving resistance data is as follows: ; in, This represents the total resistance experienced by a pure electric tractor. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates the air drag coefficient. Indicates air density, Indicates the windward area. This represents the rotational mass conversion factor. Indicates the speed of the tractor. The first derivative representing velocity, Indicates the resistance of agricultural implements during operation. This represents a coefficient related to the tractor's acceleration response. This indicates the acceleration of the tractor. Indicates the time.

[0008] Furthermore, the expression for the physical residual operator is: ; in, Represents the physical residual operator. This represents the predicted power output of the physically constrained neural network model at time t. This indicates the overall efficiency of the transmission system. Indicates the slip ratio of the drive wheel. , and These represent the static, first-order velocity, and second-order velocity shear drag coefficients, respectively. Indicates real-time tillage depth. Indicates the effective working width of the machine. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates time, Indicates the speed of the tractor. This represents the rotational mass conversion factor.

[0009] Furthermore, the expression for the comprehensive loss function is: ; ; ; ; ; in, Represents the overall objective cost function. Represents the data fitting residual term. Represents the residual term of a physical law. This represents the spatiotemporal dynamic equilibrium constraint residual term. This represents the residual term due to the running boundary constraints. , , and These represent the corresponding weight coefficients that are dynamically and adaptively adjusted based on the response characteristics of the neural tangential nucleus. This indicates the number of training samples used in a single loss calculation. Indicates the first in the batch training samples, This indicates that the physical constraint neural network model is at time 10. Output of future power demand forecasts, This represents the actual power value collected by the vehicle's sensors. Represents the square norm. Indicates the first The discrete time points corresponding to each training sample Indicates at time The difference between the predicted power and the theoretical power of the kinetic physical constraint neural network model. This represents the power demand value predicted by the physical constraint neural network model at the i-th future time step. This represents the theoretical power demand value calculated using a tractor dynamics physical constraint neural network model. Indicates the prediction time domain, This represents the decay parameter that controls the spatiotemporal dynamic balance. A time stamp representing the current moment. Indicates time, This represents the activation function of the linear rectifier unit. This represents the power predicted by the physical constraint neural network model minus the maximum power of the motor's physical limit.

[0010] Further, S5 includes: Based on the future speed state sequence and demand power prediction sequence as feedforward inputs, and using the objective cost function which includes a state of charge tracking deviation term, a control increment penalty term, a battery state of charge change rate smoothness constraint term, and a slack variable penalty term, a rolling optimization solution is performed through a physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command that satisfies the battery operating boundary and motor external characteristic constraints. Based on the optimal power allocation command, the pure electric tractor is energy managed to obtain the energy management result, thus completing the energy management of the pure electric tractor.

[0011] Furthermore, the expression for the objective cost function is: ; in, Let the quadratic objective cost function of the current control period k be represented. This represents the predicted value of the battery's state of charge. Indicates the target reference state of charge. This indicates the rate of change of battery output power. This represents the rate of change of the battery's state of charge relative to time. Indicates the prediction time domain, Indicates control of the time domain. This represents the state tracking weight matrix. This represents the control increment dynamic weight matrix. Indicates the smoothness adjustment factor. The smoothness weight matrix represents the rate of change of the state of charge. Represents the square norm. This represents the slack variables during the optimization process. Let represent the corresponding penalty coefficient, i represent the i-th future time step, and j represent the j-th future time step.

[0012] The beneficial effects of this invention are as follows: This invention provides an energy management method for pure electric tractors based on a physical constraint neural network model. By introducing a comprehensive objective cost function during the training process of the deep physical constraint neural network model, the physical constraint neural network model not only fits historical data, but also strictly follows the longitudinal dynamics of the tractor and the coupling resistance mechanism of the implement. This breaks through the black box limitation of traditional data-driven physical constraint neural network models and avoids producing non-physical prediction results that violate energy conservation under extreme variable load conditions.

[0013] When training a physical constraint neural network model, automatic differentiation technology is used to optimize parameters, which effectively avoids the problem of sensor noise being amplified through traditional discrete difference operations in the context of strong vibration in farmland, thus ensuring the accuracy of predicted input features.

[0014] The demand power prediction sequence output by the physical constraint neural network model is used as the feedforward input, and the physical constraint neural network model predictive control algorithm is used to perform rolling optimization in the prediction time domain. This mechanism enables the energy management system to predict future load changes in advance and output the optimal power allocation command under the premise of satisfying the battery operating boundary constraints and motor external characteristic constraints. It effectively avoids the large current surge in battery caused by passive following control, reduces energy consumption and extends the service life of components while ensuring operating efficiency. Attached Figure Description

[0015] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart illustrating an energy management method for a pure electric tractor based on a physically constrained neural network model, according to some embodiments of this specification. Figure 2 This is an exemplary schematic diagram of the overall architecture and information flow of the PINN-MPC energy management system for a pure electric tractor, as shown in some embodiments of this specification. Figure 3 This is an exemplary schematic diagram illustrating the longitudinal dynamics force analysis of a pure electric tractor in a complex farmland environment (including slope and implement coupling) according to some embodiments of this specification. Figure 4 This is an exemplary schematic diagram of a rolling optimization and energy optimal allocation algorithm based on physical constraint neural network model predictive control (MPC) according to some embodiments of this specification. Detailed Implementation

[0016] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0017] Example Figure 1 This is an exemplary flowchart illustrating an energy management method for a pure electric tractor based on a physically constrained neural network model, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0018] S1: Based on the multi-source sensor data collected by the pure electric tractor, feature extraction is performed using the longitudinal dynamic equation and the machine coupling resistance equation to obtain a multi-dimensional spatiotemporal feature vector.

[0019] Multi-source sensor data refers to raw information used to characterize the multi-dimensional physical operating status of pure electric tractors during complex farmland operations. For example, multi-source sensor data may include the actual speed of the drive motor, the actual ground speed of the vehicle body, the theoretical speed of the drive wheels, the slip ratio of the drive wheels, the slope of the ground undulation, the real-time traction resistance of the implements, and the current state of charge of the power battery.

[0020] In some embodiments, the processor can acquire multi-source sensor data through hardware devices such as a high-precision GPS / IMU integrated navigation unit, axle torque sensor, three-point suspension force pin sensor, and CAN bus data logger deployed on a pure electric tractor.

[0021] The longitudinal dynamics equations refer to a mathematically and physically constrained neural network model used to describe the force balance relationship of a pure electric tractor during longitudinal travel and operation. For example, the longitudinal dynamics equations may include rolling resistance to overcome ground friction, slope resistance to overcome the gravity component of the slope, acceleration inertia resistance to overcome acceleration inertia, and implement operation resistance to overcome the cutting of soil by the implements.

[0022] In some embodiments, the processor can obtain the longitudinal dynamic equations by performing kinematic and dynamic analysis on the free body of the tractor according to Newton's laws of motion. For example... Figure 3 As shown, to accurately establish a physical constraint neural network model, this invention provides a detailed decomposition of the forces acting on a pure electric tractor during field operations. The figure clearly indicates the various external forces acting on the tractor: In the vertical direction, the vehicle is subjected to its own weight G (gravity), generating a corresponding normal support force at the ground contact point, namely the front wheel's normal force F_front and the rear wheel's normal force F_rear. In the direction of travel, the tractor must overcome multiple resistances. These include: front wheel rolling resistance F_front and rear wheel rolling resistance F_rear generated by tire deformation and soil compaction; and slope resistance G (slope resistance) caused by the undulating terrain, this component of which is caused by the component of gravity along the slope direction. Furthermore, the rear wheels, as drive wheels, are also subjected to rear drive force F_rear (rear wheel driving force) provided by the drive system at the tire-ground contact surface. This force is used to balance the aforementioned travel and operational resistances, driving the tractor forward. This force analysis diagram visually illustrates the origin of each physical quantity in the longitudinal dynamic equation of the tractor, providing a clear mechanical basis for the residual constraint terms in the subsequent construction of the physical information physical constraint neural network model.

[0023] The implement coupling resistance equation refers to an empirical mathematical-physical constrained neural network model used to quantify the nonlinear load traction resistance exerted on a tractor by a suspended implement when operating in soil. For example, the implement coupling resistance equation may include static soil cutting drag coefficients, velocity-based first-order soil cutting drag coefficients, and velocity-based second-order soil cutting drag coefficients related to soil texture, as well as real-time tillage depth and effective working width of the implement.

[0024] In some embodiments, the processor can obtain the implement coupling resistance equation by invoking the ASAE (American Society of Agricultural Engineers) empirical physical constraint neural network model and combining it with a dynamic compensation mechanism based on feedback from the vehicle speed sensor.

[0025] In some embodiments, for the most typical abrupt resistance change conditions in agricultural machinery engineering (such as rotary tillage or moldboard plowing), the implement operating resistance in the implement coupling resistance physical constraint neural network model is mathematically reconstructed based on an empirical physical constraint neural network model and dynamic speed compensation is introduced. The expression for its data is: ; in, Indicates the resistance of the machinery during operation. This represents the static shear resistance coefficient associated with a specific soil texture (such as sandy loam or clay). This represents the first-order shear drag coefficient of the velocity. This represents the second-order shear drag coefficient of velocity. Indicates the speed of the tractor. This indicates the real-time tillage depth as reported by the sensor. This indicates the effective working width of the implement. This embodiment further clarifies the physical characteristic that the working resistance increases non-linearly and quadratically with the tractor speed.

[0026] Multidimensional spatiotemporal feature vectors refer to structured tensors formed by aligning and preprocessing multi-source heterogeneous sensor data with timestamps, making them suitable as inputs for deep learning physically constrained neural network models. For example, multidimensional spatiotemporal feature vectors can include normalized data such as driving speed, slip ratio, slope angle, real-time tillage depth, implement traction resistance, and battery state of charge.

[0027] In some embodiments, the processor can obtain multidimensional spatiotemporal feature vectors by performing Kalman filtering denoising and feature splicing transformation on data collected from multiple sources of sensors.

[0028] In some embodiments, the processor can collect data from multiple sources of sensors on a pure electric tractor and use the longitudinal dynamic equation, which includes rolling resistance, slope resistance, acceleration inertial resistance, and implement operation resistance, and the implement coupling resistance equation, which includes static, first-order and second-order speed soil cutting drag coefficients, to extract features and obtain a multidimensional spatiotemporal feature vector containing travel speed, slip ratio, slope angle and implement traction resistance.

[0029] In some embodiments, the expression for the total driving resistance data in the longitudinal dynamic equation is as follows: ; in, This represents the total resistance experienced by a pure electric tractor. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates the air drag coefficient. Indicates air density, Indicates the windward area. This represents the rotational mass conversion factor. Indicates the speed of the tractor. The first derivative representing velocity, Indicates the resistance of agricultural implements during operation. This represents a coefficient related to the tractor's acceleration response. This indicates the acceleration of the tractor. Indicates the time.

[0030] Figure 2 The diagram below illustrates the overall architecture and information flow of the PINN-MPC energy management system for a pure electric tractor, as provided in an embodiment of the present invention. Figure 2As shown, the overall architecture of the energy management method proposed in this invention consists of four closed-loop modules: environmental and task input, a physical information constrained neural network model power demand predictor, an energy management controller based on a physical constraint neural network model predictive control, and a pure electric tractor physical system. Specifically, in the environmental and task input stage, the system collects environmental data (including slope, road type, and load) and driving / operational requirements (including speed and power demand). This input information, combined with a predetermined path and operating mode, enters the state and information perception stage. This stage monitors in real time using sensor data (such as inertial sensors and load sensors) and vehicle status (such as speed, position, and SOC battery state of charge). The perceived state data is input to the physical information constrained neural network model power demand predictor. This predictor contains a physical constraint neural network model training and inference module and is constrained by the physical constraint neural network model during training, specifically including constraints from physical equations, dynamic equations, and the battery physical constraint neural network model. The future power sequence output by the predictor is passed as feedforward information to the energy management controller based on a physical constraint neural network model predictive control. The controller internally runs a predictive physical constraint neural network model, solving a rolling optimization algorithm (whose objective function is...). The optimization problem is addressed by strictly adhering to control objectives and constraints during the optimization process, including minimizing energy consumption, maintaining a stable State of Charge (SOC), and meeting motor / battery limitations. The optimal control sequence, calculated by the solver, is applied to the physical system of the pure electric tractor. This physical system includes the drive system (including the power system, battery pack, BMS battery management system, DC / DC converter, engine, and MCU motor controller) and a neural network model of chassis and dynamic physical constraints. The actual operating status of the system is visualized through a human-machine interface and monitoring system. The driver can view the system status (SOC, energy consumption, speed), diagnostic information, and input driver commands in real time, thus forming a closed-loop energy management and control circuit.

[0031] S2: Using a multilayer perceptron as the baseline feedforward network, a deep physical constraint neural network model architecture is constructed to obtain the physical constraint neural network model.

[0032] A physically constrained neural network model refers to a neural network model that possesses both data-driven mapping capabilities and is strictly regularized by physical laws, namely dynamic equilibrium and energy conservation. For example, the structure of a physically constrained neural network model may include an input layer, multiple hidden layers, and an output layer. The input layer receives the multidimensional spatiotemporal feature vector and performs dimension matching processing; multiple hidden layers are used to perform feedforward nonlinear mapping processing on the multidimensional spatiotemporal feature vector. In the construction of the computational graph of the hidden layers, in addition to establishing power mapping nodes, a speed prediction auxiliary branch for characterizing the longitudinal motion state of the vehicle is simultaneously established. This auxiliary branch, based on the multidimensional spatiotemporal feature vector and continuous time variables, outputs continuous speed state prediction values ​​through hidden layer neurons. The activation function of the hidden layer uses a hyperbolic tangent function or a sine function with continuous derivatives at zero to ensure that the speed prediction branch is second-order continuously differentiable with respect to the time variable, thus obtaining the hidden layer features; the output layer is used to transform the hidden layer features into prediction results and output the future speed state sequence and the future demand power sequence.

[0033] A future speed state sequence refers to a continuous or discrete prediction array of a tractor's speed over a finite future time span, generated by a physically constrained neural network model in its hidden or auxiliary output layers. For example, a future speed state sequence could include the theoretical expected vehicle speed at 30 consecutive discrete time points within the next 3 seconds, divided into steps of 100ms. This sequence functions as a continuous-time variable within the network and directly participates in automatic differentiation calculations.

[0034] In some embodiments, the processor can obtain the future velocity state sequence by using the hidden layer velocity prediction branch of the physically constrained neural network model, and perform forward calculation using multidimensional spatiotemporal feature vectors and historical state sequences. Based on this sequence node, the processor can directly calculate the acceleration term by using the underlying automatic differentiation mechanism to differentiate its continuous time variables.

[0035] In some embodiments, the processor can obtain a physically constrained neural network model by instantiating the network parameter matrix of a multilayer perceptron (MLP) in memory and loading initial weights.

[0036] S3: Based on multidimensional spatiotemporal feature vectors, historical state sequences, and sensor measured power data, combined with longitudinal dynamic equations, and using a comprehensive loss function, the physical constraint neural network model is trained through backpropagation and automatic differentiation techniques to obtain a trained physical constraint neural network model.

[0037] A historical state sequence refers to a time-series data set of the continuous operating status of a tractor within a sliding time window preceding the current time point. For example, a historical state sequence may include speed change curves, slip ratio fluctuation records, and historical time-series values ​​of implement resistance over the past 3 to 5 seconds.

[0038] In some embodiments, the processor can obtain historical state sequences through the vehicle CAN bus cache queue and memory sliding window extraction mechanism.

[0039] Sensor-measured power data refers to the output power label data of the drive system directly recorded by measuring equipment during the actual operation of the tractor. For example, sensor-measured power data may include the bus discharge power calculated by multiplying the voltage sensor and the current sensor, or the mechanical output power obtained by multiplying the wheel-side torque sensor and the speed sensor.

[0040] In some embodiments, the processor can obtain sensor measured power data by reading the historical operation log file of the vehicle controller.

[0041] The comprehensive loss function is a cost evaluation function used in the backpropagation training of a physically constrained neural network model to measure the total deviation between the model's output and multiple optimization objectives. For example, the comprehensive loss function may include a data fitting residual term that measures the deviation between the predicted value and the label, a physical law residual term that measures whether physical laws are violated, a spatiotemporal dynamic equilibrium constraint residual term that measures the continuity of future power, and an operational boundary constraint residual term that measures whether the limits are exceeded.

[0042] In some embodiments, the processor can obtain the comprehensive loss function by calling the tensor computation interface in the deep learning framework to sum the individual residual components with adaptive weights.

[0043] In some embodiments, the processor can extract the core resistance component of the total driving resistance data based on multi-dimensional spatiotemporal feature vectors, historical state sequences, and sensor measured power data, combined with the longitudinal dynamic equation, and construct a physical residual operator; use an automatic differentiation mechanism to construct a computational graph to directly differentiate the future velocity state sequence output by the physical constraint neural network model with respect to continuous time variables to calculate the acceleration term; use the physical residual operator and the acceleration term to calculate the physical law residual term; use a comprehensive loss function containing data fitting residual term, physical law residual term, spatiotemporal dynamic equilibrium constraint residual term, and running boundary constraint residual term to train the physical constraint neural network model through backpropagation and automatic differentiation techniques to obtain a trained physical constraint neural network model.

[0044] The physical residual operator refers to the difference term generated when the predicted output of a physically constrained neural network model is substituted into the physical conservation equations, due to the non-closure of the theoretical physical equations caused by errors in the physically constrained neural network model. For example, the physical residual operator can include the difference between the power output predicted by the physically constrained neural network model and the theoretical power demand derived from the tractor's travel resistance and the machine's operating resistance.

[0045] In some embodiments, the processor can obtain the physical residual operator by performing a difference operation between the real-time prediction results of the physical constraint neural network model and the theoretical calculation results based on the longitudinal dynamic equation.

[0046] The acceleration term is a physical quantity that characterizes how quickly the tractor's speed changes over time. It is used to calculate the additional resistance required to overcome the system's inertia during vehicle acceleration. For example, the acceleration term may include the first derivative of velocity, used to calculate conventional acceleration resistance, and the second derivative, used to characterize the rate of change of acceleration and transient shocks.

[0047] In some embodiments, the processor can obtain the acceleration term by directly performing implicit differentiation on the input continuous-time variable neurons by calling the automatic differential computation graph built into the deep learning framework, thus avoiding the amplification of noise by discrete difference.

[0048] In some embodiments, to avoid sensor noise being amplified exponentially through discrete difference operations in strong vibration environments in farmland, the physical constraint neural network model abandons the traditional difference formula when calculating the acceleration term. Instead, it employs a low-level automatic differentiation (AD) mechanism to directly differentiate continuous-time variables. The expression for its automatic differentiation calculation is as follows: ; in, This indicates the tractor's speed relative to time. The exact first derivative (i.e., theoretical acceleration); This represents the time-related information constructed when a physically constrained neural network model passes input features to deeper layers. The speed output node function; This represents the partial derivative obtained by directly differentiating the velocity output node function using the chain rule through the computational graph mechanism built into the deep learning framework (such as inverse mode automatic differentiation). Through this mechanism, the system can obtain a precise acceleration term unaffected by high-frequency mechanical noise without relying on the discrete sampling differences between adjacent sensor time points.

[0049] In some embodiments, a training process for a physically constrained neural network model is provided: The processor first acquires multi-dimensional spatiotemporal feature vectors and historical state sequences as training input data, and acquires corresponding sensor measured power data as labels. The training data is input into the initialized physically constrained neural network model, which performs forward propagation processing through its input layer and hidden layers, obtaining the initial demand power prediction result at the output layer. Subsequently, the error calculation stage begins: The processor extracts the data fitting residual term from the initial prediction result and the sensor measured power data; simultaneously, it constructs a computational graph using an automatic differentiation mechanism, directly differentiating the future velocity state sequence output by the network with respect to continuous time variables to obtain the acceleration term, substituting this acceleration term and the prediction result into the longitudinal dynamic equation to construct a theoretically non-closed physical residual operator, and then calculating the physical law residual term; then, combining the power continuity in the future time domain and the maximum power extremum of the motor, it calculates the spatiotemporal dynamic equilibrium constraint residual term and the operating boundary constraint residual term respectively. The processor adds the above four terms according to the dynamic adaptive weight coefficients to obtain the comprehensive loss function. Finally, based on the comprehensive loss function, the processor calculates the gradient of the network weights using the backpropagation algorithm, and uses an optimizer (such as Adam) to update and tune the parameters in the solution space. The above forward and backward propagation processes are iterated until the value of the comprehensive loss function converges to the preset error threshold or the maximum number of iterations is reached, training is complete, and the trained physical constraint neural network model is finally obtained.

[0050] In some embodiments, the expression for the physical residual operator is: ; in, Represents the physical residual operator. This represents the predicted power output of the physically constrained neural network model at time t. This indicates the overall efficiency of the transmission system. Indicates the slip ratio of the drive wheel. , and These represent the static, first-order velocity, and second-order velocity shear drag coefficients, respectively. Indicates real-time tillage depth. Indicates the effective working width of the machine. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates time, Indicates the speed of the tractor. This represents the rotational mass conversion factor.

[0051] In some embodiments, the expression for the comprehensive loss function is: ; ; ; ; ; in, Represents the overall objective cost function. Represents the data fitting residual term. Represents the residual term of a physical law. This represents the spatiotemporal dynamic equilibrium constraint residual term. This represents the residual term due to the running boundary constraints. , , and These represent the corresponding weight coefficients that are dynamically and adaptively adjusted based on the response characteristics of the neural tangential nucleus. This indicates the number of training samples used in a single loss calculation. Indicates the first in the batch training samples, This indicates that the physical constraint neural network model is at time 10. Output of future power demand forecasts, This represents the actual power value collected by the vehicle's sensors. Represents the square norm. Indicates the first The discrete time points corresponding to each training sample Indicates at time The difference between the predicted power and the theoretical power of the kinetic physical constraint neural network model. This represents the power demand value predicted by the physical constraint neural network model at the i-th future time step. This represents the theoretical power demand value calculated using a tractor dynamics physical constraint neural network model. Indicates the prediction time domain, This represents the decay parameter that controls the spatiotemporal dynamic balance. A time stamp representing the current moment. Indicates time, This represents the activation function of the linear rectifier unit. This represents the power predicted by the physical constraint neural network model minus the maximum power of the motor's physical limit.

[0052] In some embodiments, to address the significant differences in numerical magnitude, gradient sensitivity, and convergence speed among the residuals in the comprehensive loss function and to avoid gradient imbalance during network training, the physically constrained neural network model employs a dynamic adaptive annealing adjustment strategy based on the response characteristics of the Neural Tangent Kernel (NTK) to update the weight coefficients during parameter optimization. Specifically, in the early stages of training, the processor increases the weight coefficients of the terms representing the data fitting residuals. The proportion of this allows the network to prioritize learning the distribution patterns of sensor measured power data; as training iteratively progresses, the weight coefficients representing the residual terms of physical laws are gradually increased. The weighting coefficients representing the spatiotemporal dynamic equilibrium constraint residuals The proportion of this factor enables the physical constraint neural network model to converge in a direction that satisfies the longitudinal dynamic constraints of the tractor and the consistency of future predicted time-domain power evolution; simultaneously, the weight coefficients representing the residual terms of the operational boundary constraints are enhanced in the later stages of training. The constraint effect is used to suppress unreasonable outputs where the predicted power exceeds the physical limits of the motor. Through this strategy, the physically constrained neural network model can adaptively balance the relationship between data-driven learning and the embedding of physical mechanisms at different training stages.

[0053] S4: Based on the real-time acquired state sequence of the pure electric tractor, the trained physical constraint neural network model is used to calculate the future speed state sequence and the future demand power prediction sequence within the finite time domain.

[0054] A pure electric tractor state sequence refers to the latest operating condition data stream continuously generated by a sensor array during the tractor's online real-time operation. For example, a pure electric tractor state sequence may include the instantaneous travel speed, instantaneous gradient, instantaneous slip ratio, and instantaneous tool resistance at the current time t.

[0055] In some embodiments, the processor can obtain the state sequence of a pure electric tractor by frequently polling the on-board sensor bus interface and performing real-time filtering and noise reduction.

[0056] A demand power prediction sequence is a discrete prediction array of the driving power required by a tractor system over a finite future time span, output by a trained physical constraint neural network model. For example, a demand power prediction sequence may include high-precision expected power values ​​at 30 consecutive discrete time points within the next 3 seconds, divided into steps of 100ms.

[0057] In some embodiments, the processor can obtain a demand power prediction sequence by inputting the state sequence of the pure electric tractor into a trained physical constraint neural network model for forward inference calculation.

[0058] In some embodiments, the process of obtaining the demand power prediction sequence based on the real-time acquired state sequence of a pure electric tractor and using a trained physical constraint neural network model can be achieved through the following nonlinear mapping expression: ; in, Indicates the current time And the predicted power output in the future time domain; This represents the proxy physical constraint neural network model function after the physical constraint neural network model has converged through backpropagation and automatic differentiation techniques. Indicates the instantaneous speed of the tractor; Indicates the slip ratio of the drive wheel; Indicates the current slope angle of the farmland; Indicates the real-time traction resistance of the work equipment; This represents the current state of charge of the power battery. This expression demonstrates that the trained physical constraint neural network model can directly receive multi-dimensional heterogeneous operating condition feature inputs and, during forward propagation, directly output a high-precision demand power prediction sequence based on learned historical distribution patterns and internalized dynamic mechanisms.

[0059] S5: Based on the future speed state sequence and demand power prediction sequence, the rolling optimization solution is performed using the physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command for the pure electric tractor. Energy management is then performed on the pure electric tractor according to the optimal power allocation command to obtain the energy management result and complete the energy management of the pure electric tractor.

[0060] Physically constrained neural network model predictive control (PCM) refers to an advanced control strategy based on a physically constrained neural network model of a system, which predicts future system behavior and performs rolling optimization within a finite time domain. For example, a PCM model predictive control algorithm may include a state-space physically constrained neural network model construction module, a future state prediction calculation module, and a quadratic programming solution module.

[0061] In some embodiments, the processor can obtain a physically constrained neural network model predictive control algorithm by loading a discretized power battery state-space equation into the energy management controller and configuring rolling time-domain parameters.

[0062] The optimal power allocation command refers to the control signal that the system determines, after optimization, to be the most energy-efficient and safest under the current operating conditions and in the future prediction time domain. For example, the optimal power allocation command may include the expected discharge power value issued to the battery management system (BMS) and the target torque request value issued to the motor inverter.

[0063] In some embodiments, the processor can obtain the optimal power allocation instruction by solving the objective cost function using quadratic programming (QP) and extracting the first control increment at the current moment.

[0064] Energy management results refer to the objective physical characteristics of the vehicle's powertrain system in terms of energy distribution, consumption, and component health status after the system executes actions according to optimal instructions. For example, energy management results may include specific data on the suppression of transient voltage drops at the bus, the reduction in battery charging and discharging current fluctuations, and the reduction in overall energy consumption.

[0065] In some embodiments, the processor can obtain energy management results by continuously monitoring the actual feedback signals of the underlying actuators and the actual change trajectory of the battery SOC.

[0066] In some embodiments, the processor can use the demand power prediction sequence as a feedforward input and a target cost function that includes a state of charge tracking deviation term, a control increment penalty term, a battery state of charge change rate smoothness constraint term, and a slack variable penalty term. The processor can then perform rolling optimization by using a physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command that satisfies the battery operating boundary and motor external characteristic constraints. Based on the optimal power allocation command, the processor can perform energy management on the pure electric tractor to obtain the energy management result and complete the energy management of the pure electric tractor.

[0067] The objective cost function is a quadratic mathematical expression used to quantify the performance of system control and to minimize the objective in the rolling optimization process of physical constraint neural network model predictive control. For example, the objective cost function may include a state-of-charge tracking deviation term to constrain the battery within a safe range, a control increment penalty term to suppress current step shocks, a smoothness constraint term for the rate of change of battery state of charge to mitigate battery aging, and a slack variable penalty term to ensure the feasibility of the solution.

[0068] In some embodiments, the processor can obtain the target cost function by allocating different weight matrices (Q, R, S matrices) according to the emphasis of the energy management strategy and combining them with the system state equations to perform algebraic construction.

[0069] In some embodiments, the expression for the objective cost function is: ; in, Let the quadratic objective cost function of the current control period k be represented. This represents the predicted value of the battery's state of charge. Indicates the target reference state of charge. This indicates the rate of change of battery output power. This represents the rate of change of the battery's state of charge relative to time. Indicates the prediction time domain, Indicates control of the time domain. This represents the state tracking weight matrix. This represents the control increment dynamic weight matrix. Indicates the smoothness adjustment factor. The smoothness weight matrix represents the rate of change of the state of charge. Represents the square norm. This represents the slack variables during the optimization process. Let represent the corresponding penalty coefficient, i represent the i-th future time step, and j represent the j-th future time step.

[0070] In some embodiments, before performing rolling optimization using a physical constraint neural network model predictive control algorithm, it is necessary to construct a state-space physical constraint neural network model describing the dynamic behavior of the system. For the energy dispatch of a power battery, the data expression of its discrete state-space equations is as follows: ; in, Indicating the future The predicted state of charge (SOC) of the power battery in the current control cycle; $SOC(k)$ represents the predicted SOC value in the current control cycle. The initial state of charge of the power battery in each control cycle; Indicates the first The desired output power command of the battery in each control cycle (i.e., the system control variable). This represents the discrete control step size for predictive control using a physical constraint neural network model. This represents the bus voltage at the current moment; This represents the rated capacity of the power battery; the constant 3600 is used to convert hours to seconds to unify the physical dimensions. This state-space equation provides the fundamental mathematical constraints for the rolling optimization of the objective cost function, illustrating the evolution of the system state over time.

[0071] In some embodiments, the process of obtaining the optimal power allocation command through rolling optimization using a physical constraint neural network model predictive control algorithm is limited by the actual physical capabilities of the vehicle and must satisfy multiple hard boundary constraints. These include: (1) Power battery output power constraint: ,in Indicates at any control moment The expected output power of the battery, and These represent the minimum and maximum allowable discharge power limits of the battery, respectively, to prevent overcharging or high-rate discharge.

[0072] (2) External characteristic torque constraint of the drive motor: ,in This indicates the output torque of the motor. Indicates the current speed The maximum external characteristic envelope torque of the motor is set to ensure that the motor is not overloaded.

[0073] (3) Battery state of charge safe operating range constraints: ,in It represents the state of charge in the future predicted time domain, in order to balance the depth of discharge control and lifetime maintenance.

[0074] (4) Constraints of the vehicle energy flow power balance equation: ,in This represents the battery output power at the current moment. For DC-side energy transfer efficiency. This is the predicted power demand value at the current moment. To mitigate the additional power consumption of the auxiliary systems and ensure energy conservation in scheduling.

[0075] In some embodiments, such as Figure 4 As shown, the working mechanism of the MPC controller is a closed-loop iterative process, including the following key steps: First, the MPC controller receives input signals, the core of which lies in its built-in predictive physical constraint neural network model. This physical constraint neural network model infers the expected future dynamic behavior based on the current state of the controlled object. During this process, the rolling optimization module, combined with the reference trajectory provided to the controller (i.e., the desired SOC target, etc.), uses the solver to calculate a set of optimal control sequences under the premise of satisfying system and constraint conditions (such as motor torque limit, battery power limit). In the specific implementation mechanism, the controller does not execute the entire set of optimal sequences, but only executes the first control variable, issuing the control command to the controlled object (i.e., the pure electric tractor drive system). Subsequently, the controlled object executes the command and generates actual system outputs (such as speed changes, SOC consumption). At this time, the state estimator collects the current state information of the system and feeds it back to the controller. The controller then moves to the next moment, combining the updated state information to continuously update the predictive physical constraint neural network model and the optimization problem, starting a new round of solving and execution. This cyclical mechanism of rolling optimization and point-by-point implementation ensures that the system can always achieve the global optimal allocation of energy flow based on the latest vehicle status and prediction information when facing complex farmland variable load conditions.

Claims

1. An energy management method for pure electric tractors based on a physically constrained neural network model, characterized in that, include: S1: Based on the multi-source sensor data collected by the pure electric tractor, feature extraction is performed using the longitudinal dynamic equation and the machine coupling resistance equation to obtain a multi-dimensional spatiotemporal feature vector; S2: Using a multilayer perceptron as the baseline feedforward network, a deep physical constraint neural network model architecture is constructed to obtain the physical constraint neural network model; S3: Based on multidimensional spatiotemporal feature vectors, historical state sequences and sensor measured power data, combined with longitudinal dynamic equations, and using a comprehensive loss function, the physical constraint neural network model is trained through backpropagation and automatic differentiation techniques to obtain a trained physical constraint neural network model. S4: Based on the real-time acquired state sequence of the pure electric tractor, the trained physical constraint neural network model is used to calculate the future speed state sequence and the future demand power prediction sequence within the finite time domain. S5: Based on the future speed state sequence and demand power prediction sequence, the rolling optimization solution is performed using the physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command for the pure electric tractor. Energy management is then performed on the pure electric tractor according to the optimal power allocation command to obtain the energy management result and complete the energy management of the pure electric tractor.

2. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 1, characterized in that, S1 includes: Based on multi-source sensor data collected by pure electric tractors, the longitudinal dynamic equation, which includes rolling resistance, slope resistance, acceleration inertial resistance, and implement operation resistance, is used to extract features from the implement coupling resistance equation, which includes static, first-order and second-order speed soil cutting drag coefficients. This yields a multi-dimensional spatiotemporal feature vector containing travel speed, slip ratio, slope angle, and implement traction resistance.

3. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 1, characterized in that, S3 includes: Based on multidimensional spatiotemporal feature vectors, historical state sequences and sensor measured power data, the core resistance component of total driving resistance data is extracted by combining longitudinal dynamic equations, and a physical residual operator is constructed. The acceleration term is calculated by directly differentiating the future velocity state sequence output by the physical constraint neural network model with respect to continuous time variables using an automatic differentiation mechanism to construct a computational graph. The physical law residual terms are obtained by using the physical residual operator and the acceleration term; By using a comprehensive loss function that includes data fitting residuals, physical law residuals, spatiotemporal dynamic equilibrium constraint residuals, and operational boundary constraint residuals, a physical constraint neural network model is trained using backpropagation and automatic differentiation techniques, resulting in a well-trained physical constraint neural network model.

4. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 3, characterized in that, In the longitudinal dynamic equation, the expression for the total driving resistance data is: ； in, This represents the total resistance experienced by a pure electric tractor. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates the air drag coefficient. Indicates air density, Indicates the windward area. This represents the rotational mass conversion factor. Indicates the speed of the tractor. The first derivative representing velocity, Indicates the resistance of agricultural implements during operation. This represents a coefficient related to the tractor's acceleration response. This indicates the acceleration of the tractor. Indicates the time.

5. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 3, characterized in that, The expression for the physical residual operator is: ； in, Represents the physical residual operator. This represents the predicted power output of the physically constrained neural network model at time t. This indicates the overall efficiency of the transmission system. Indicates the slip ratio of the drive wheel. , and These represent the static, first-order velocity, and second-order velocity shear drag coefficients, respectively. Indicates real-time tillage depth. Indicates the effective working width of the machine. This indicates the total mass of the tractor and implements. Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the current slope angle of the farmland. Indicates time, Indicates the speed of the tractor. This represents the rotational mass conversion factor.

6. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 3, characterized in that, The expression for the comprehensive loss function is: ； ； ； ； ； in, Represents the overall objective cost function. Represents the data fitting residual term. Represents the residual term of a physical law. This represents the spatiotemporal dynamic equilibrium constraint residual term. This represents the residual term due to the running boundary constraints. , , and These represent the corresponding weight coefficients that are dynamically and adaptively adjusted based on the response characteristics of the neural tangential nucleus. This indicates the number of training samples used in a single loss calculation. Indicates the first in the batch training samples, This indicates that the physical constraint neural network model is at time 10. Output of future power demand forecasts, This represents the actual power value collected by the vehicle's sensors. Denotes the square norm. Indicates the first The discrete time points corresponding to each training sample Indicates at time The difference between the predicted power and the theoretical power of the kinetic physical constraint neural network model. This represents the power demand value predicted by the physical constraint neural network model at the i-th future time step. This represents the theoretical power demand value calculated using a tractor dynamics physical constraint neural network model. Indicates the prediction time domain, This represents the decay parameter that controls the spatiotemporal dynamic balance. A time stamp representing the current moment. Indicates time, This represents the activation function of the linear rectifier unit. This represents the power predicted by the physical constraint neural network model minus the maximum power of the motor's physical limit.

7. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 1, characterized in that, S5 includes: Based on the future speed state sequence and demand power prediction sequence as feedforward inputs, and using the objective cost function which includes a state of charge tracking deviation term, a control increment penalty term, a battery state of charge change rate smoothness constraint term, and a slack variable penalty term, a rolling optimization solution is performed through a physical constraint neural network model predictive control algorithm to obtain the optimal power allocation command that satisfies the battery operating boundary and motor external characteristic constraints. Based on the optimal power allocation command, the pure electric tractor is energy managed to obtain the energy management result, thus completing the energy management of the pure electric tractor.

8. The energy management method for pure electric tractors based on a physically constrained neural network model according to claim 7, characterized in that, The expression for the objective cost function is: ； in, Let the quadratic objective cost function of the current control period k be represented. This represents the predicted value of the battery's state of charge. Indicates the target reference state of charge. This indicates the rate of change of battery output power. This represents the rate of change of the battery's state of charge relative to time. Indicates the prediction time domain, Indicates control of the time domain. This represents the state tracking weight matrix. This represents the control increment dynamic weight matrix. Indicates the smoothness adjustment factor. The smoothness weight matrix represents the rate of change of the state of charge. Denotes the square norm. This represents the slack variables during the optimization process. Let represent the corresponding penalty coefficient, i represent the i-th future time step, and j represent the j-th future time step.

Images