Method for short-time load prediction based on igwo-tsmixer rotary tillage operation battery power distribution

Abstract

The present application relates to the technical field of electric agricultural equipment energy management, and specifically discloses a rotary tillage operation battery power distribution method based on IGWO-TSMixer short-time load prediction, which comprises the following steps: collecting multi-source time sequence data of driving speed, actual plowing depth and PTO torque in the rotary tillage operation within 2-5s of history, constructing a sliding time window sample set, and dividing the sample set into a training set, a verification set and a test set in time sequence; the present application constructs a TSMixer multivariate time sequence prediction model based on a pure MLP architecture, extracts time sequence dependence and variable interaction in the historical load sequence by combining time mixing and feature mixing double modules, realizes direct multi-step prediction of future short-time traction resistance and rotary tillage load, introduces an improved grey wolf optimization algorithm to perform adaptive optimization on key hyperparameters of the TSMixer, so that the model can maintain stable prediction accuracy under different plowing depths and load fluctuations, and significantly enhances the adaptability of the prediction model to different soils and operation conditions.

Description

Technical Field

[0001] This invention relates to the field of energy management technology for electric agricultural equipment, and in particular to a method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer. Background Technology

[0002] Rotary tillage is one of the most dramatic load fluctuations in the operation of electric horticultural multi-functional platforms. During operation, the platform needs to simultaneously drive the tracked chassis to overcome soil resistance to maintain a stable travel speed and drive the rotary tiller shaft to complete soil cutting, crushing, and throwing operations. Its power demand consists of two parts: the walking traction load and the rotary tillage load. The continuous changes in factors such as soil firmness, actual tillage depth, surface unevenness, operating speed, and the load state of the tiller shaft cause both traction resistance and rotary tillage resistance to exhibit obvious non-stationarity and short-term random fluctuation characteristics. Under such dynamic operating conditions, the battery system needs to simultaneously meet the power output requirements of walking drive and rotary tillage operations. Its output power is limited by constraints such as battery state of charge, maximum output power, current limit, and power change rate. The rationality of power distribution has a direct impact on the overall machine's operational stability and energy efficiency.

[0003] In existing rotary tillage operations, traction resistance and PTO (power output shaft) torque fluctuate drastically with soil firmness, tillage depth, and speed. The battery passively supplies power according to the current load, which leads to problems such as power response lag, increased current surge, and decreased motor speed. Moreover, the hyperparameters of the TSMixer (time series mixer) model are often tuned empirically, resulting in significant fluctuations in prediction errors under different tillage depths. The generalization performance is difficult to stably adapt to changing working conditions. Furthermore, when the predicted load exceeds the battery's available power, there is a lack of a forward-looking power coordination mechanism. The power distribution between the walking and rotary tillage systems conflicts under continuous heavy loads, making it difficult to balance driving stability and work quality. Therefore, how to solve the problems of accurate short-term load prediction, adaptive parameter tuning of the model under working conditions, and coordinated power distribution under heavy loads is the technical problem that this invention aims to solve. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a battery power allocation method for rotary tillage operations based on short-term load prediction using IGWO-TSMixer, which can effectively solve the problems mentioned in the prior art.

[0005] The objective of this invention can be achieved through the following technical solution: This invention provides a battery power allocation method for rotary tillage operations based on short-term load prediction using IGWO-TSMixer, the specific steps of which are as follows: Step 1: Collect multi-source time-series data on driving speed, actual tillage depth, and PTO torque within 2-5 seconds of historical data during rotary tillage operations, construct a sliding time window sample set, and divide it into training set, validation set, and test set according to time order to build a standardized sample library covering multiple working conditions, laying the data foundation for subsequent model training and generalization evaluation. Step 2: Construct a TSMixer multivariate time-series prediction model based on a pure MLP architecture. Extract time-series dependencies and variable interaction relationships through time mixing and feature mixing modules. Directly predict traction resistance and rotary tillage load in the next 2-5 seconds in multiple steps to obtain future load trends in advance. TSMixer is a time series mixer that realizes lightweight multi-step prediction and provides real-time and reliable load change trend information for power allocation. Step 3: Introduce the improved gray wolf optimization algorithm (IGWO), which encodes the learning rate, dropout coefficient and hidden layer dimension of TSMixer as gray wolf positions, and uses the root mean square error of the validation set as the fitness to adaptively optimize the hyperparameters. The improved gray wolf optimization algorithm is IGWO, which enhances the generalization stability under different tillage depth conditions, reduces the cost of manual parameter tuning and improves the consistency of prediction across working conditions. Step 4: Utilize the optimized IGWO-TSMixer to output the predicted future load value, combine the traction resistance and PTO torque to estimate the future power demand of the walking and rotary tillage systems, and calculate the current available power based on the battery SOC and current limit to determine whether the system has entered a heavy load state. This will help identify the power supply and demand contradiction in advance and provide a reliable triggering basis for active coordination control under heavy load conditions. Step 5: Based on the judgment results, implement a graded power coordination strategy: smooth the output under normal load, dynamically adjust the priority coefficients of walking and rotary tillage under continuous heavy load, and reduce the target driving speed in advance to achieve stable battery power distribution and work quality optimization, realize the coordinated control of power smoothness and work stability, and take into account both battery safety and rotary tillage work quality.

[0006] Preferably, step 1 specifically includes: The electric gardening operation platform is synchronously collected at a fixed sampling period, including its driving speed, actual tillage depth, platform roll and pitch angles, traction resistance, PTO torque, rotary tillage motor power, battery voltage, battery current, and battery power under rotary tillage conditions. This constitutes a multi-source time-series data matrix, enabling the coordinated acquisition of multiple physical quantities and providing a complete state information basis for load prediction. A sliding time window is used to construct supervised learning samples. The length of the historical input window is set to 2-5s, and the length of the future prediction window is set to 2-5s. The multi-source data of the historical window corresponding to each sampling time is used as the input features, and the traction resistance, PTO torque and rotary tillage power sequence in the future window are used as prediction labels to generate a sample set, which supports the model to learn the temporal causal relationship and improves the prediction reliability. The sample set is divided into 70% training set, 15% validation set, and 15% test set in chronological order to ensure that data within the same time window is not distributed across sets. The statistics of the training set are calculated separately and used for normalization processing of all subsequent data groups to avoid leakage of time series information, ensure the independence and objectivity of each stage, avoid data crossing causing inflated performance, and ensure fair evaluation.

[0007] Preferably, step 2 specifically includes: A TSMixer multivariate time series prediction model based on a pure MLP architecture is constructed. The multi-source load sequences of the historical input window are simultaneously fed into the time mixing module and the feature mixing module. Through the fully connected mapping across time steps, the short-term trend and periodic fluctuation dependencies between different historical moments are extracted, accurately capturing the short-term fluctuation patterns of the load and providing reliable time series features for power feedforward control. The intermediate feature sequence output by the time mixing module is fed into the feature mixing module, and a fully connected mapping is performed along the feature dimension to extract the variable interaction relationship between driving speed, actual tillage depth, traction resistance, PTO torque and battery state variables. The information transmission efficiency is enhanced by residual connection and layer normalization, effectively integrating multi-variable coupled information and improving the model's comprehensive perception ability of complex working conditions. Multiple TSMixer modules are stacked to construct a deep prediction network. The final encoded time-series features are mapped by a linear prediction head to the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence for each prediction step within the next 2 to 5 seconds. This enables direct multi-step output of multivariate time-series data, outputting multi-step prediction results at once, avoiding the accumulation of recursive errors, and improving prediction efficiency and stability.

[0008] Preferably, step 2 further includes: The multi-source time series data matrix input to the current TSMixer module is normalized by the layer and then transposed along the time dimension. After transposition, the time step dimension of the matrix is ​​transformed into the feature dimension, and the feature dimension is transformed into the time step dimension. The transformation result is input into the time mixing fully connected layer for fully connected mapping between time steps. The mapping result is reversed and restored, and then superimposed with the original multi-source time series data matrix through residual connection to obtain the time mixing output, which enhances the efficiency of time series feature transfer and improves the modeling accuracy of dependencies between historical time points. After the time-mixed output is normalized by the layer, it is input into the feature mixing fully connected layer along the feature dimension to perform fully connected mapping between variables. This extracts the cross-variable coupling relationship between driving speed, actual tillage depth, traction resistance, PTO torque and battery state variables. The mapping result is superimposed by residual connection to obtain the feature mixing output, which completes the information integration of a single TSMixer module, strengthens the collaborative expression of multiple variables, and improves the model's overall perception ability of complex working conditions. Multiple TSMixer modules are stacked sequentially, with the feature sequence output by the previous module serving as the input sequence for the next module. Through progressively layered time mixing and feature mixing, a multi-level abstract representation of historical load sequences, from short-term fluctuations to long-term trends, is constructed to obtain future load trends in advance. Finally, the fully connected prediction head directly maps the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence for each prediction step within the next 2 to 5 seconds, abstracting time-series features layer by layer and enhancing the model's comprehensive ability to express the laws of load evolution.

[0009] Preferably, the TSMixer multivariate time series prediction model adopts a pure MLP architecture to replace the recursive and attention structure. Its time mixing and feature mixing dual branches are both based on fully connected layers to achieve positive mapping, which simplifies the calculation process, reduces the deployment difficulty, and improves the real-time prediction response speed. Moreover, the time mixing and feature mixing dual branches perform cross-parameter full connection along the variable dimension, directly modeling the instantaneous coupling relationship between speed-resistance-torque-power, strengthening the ability to express the instantaneous correlation between physical variables, and improving the consistency of multi-objective prediction.

[0010] Preferably, step 3 specifically includes: The initial learning rate, L2 regularization coefficient, Dropout coefficient, number of module stacking layers, temporal hybrid hidden dimension, and feature hybrid hidden dimension of the TSMixer multivariate temporal prediction model are encoded into the position vectors of individual gray wolves in the improved gray wolf optimization algorithm. The search boundary range of each hyperparameter is set, and the hyperparameter combination is encoded into the position vector to ensure that the search boundary is reasonable, realize the systematic search of key hyperparameters, and avoid the blindness of manual trial and error. The weighted average of the three errors of traction resistance prediction RMSE, PTO torque prediction RMSE and rotary tillage power prediction RMSE on the validation set is used as the fitness function. A cosine adaptive convergence factor is used to dynamically adjust the intensity of global exploration and local development. A particle swarm velocity memory mechanism is introduced to retain the historical search direction, dynamically balance global exploration and local development, improve search efficiency, retain the historical search direction, accelerate convergence and improve stability. A random difference mutation strategy is applied to the current optimal gray wolf position to generate new candidate solutions. Individuals with better fitness are retained through greedy selection. The gray wolf population is iteratively updated until the maximum number of generations is reached. The optimal hyperparameter combination that minimizes the multi-task prediction error on the validation set is output, which enhances population diversity, suppresses premature convergence, outputs the global optimal solution, and improves the generalization ability of cross-condition prediction.

[0011] Preferably, step 4 specifically includes: The real-time collected historical multi-source data is input into the optimized IGWO-TSMixer prediction model. The model directly outputs the predicted values ​​of traction resistance, PTO torque and rotary tillage power for each prediction step in the next 2 to 5 seconds. This serves as the information feedforward basis for forward-looking power allocation, enabling the early acquisition of load trends, providing time margin for power allocation, reducing the risk of response lag, and improving the proactive control capability. The future driving power of the walking system is estimated based on the predicted traction resistance and the current driving speed. The future operating power of the rotary tillage system is estimated based on the predicted PTO torque and the cutter shaft angular velocity. The sum of the two and the power of the auxiliary system are added to obtain the future total battery power demand sequence. The power demand is quantified in advance, providing clear indicators for battery scheduling and improving the foresight and accuracy of energy management. The system collects the current battery SOC, terminal voltage, and bus current. Based on the SOC correction factor and the maximum allowable discharge current, it calculates the current available power of the battery. It then compares the future total power demand sequence with the available power step by step to determine whether each prediction step is under heavy load. The system assesses the power margin in real time, accurately identifies overload risks, and provides a reliable triggering basis for graded power coordination.

[0012] Preferably, step 4 further includes: Set the battery's safe operating range and maximum allowable discharge current. The battery's available power is calculated by multiplying the current terminal voltage, the maximum allowable current, and the SOC correction factor. When the SOC is below the lower threshold, the correction factor is linearly reduced to limit the discharge rate. The upper limit of available power is dynamically adjusted to prevent over-discharge at low SOC from damaging the battery's lifespan. When estimating the power of the walking drive system, the efficiency of the walking drive system is taken into account. The power is obtained by dividing the product of the predicted traction resistance and the predicted travel speed by the efficiency. When estimating the power of the rotary tillage operation, the transmission efficiency of the rotary tillage operation is taken into account. The power is obtained by dividing the product of the predicted PTO torque and the predicted cutter shaft angular velocity by the efficiency. Transmission loss correction is introduced to improve the actual consistency and reliability of the power estimation. When the total power demand in any prediction step within the future prediction time domain exceeds the available battery power, the prediction window is marked as an overload state, and the overload duration and maximum power excess are recorded. This provides a basis for tiered decision-making for subsequent power coordination strategies, accurately quantifies the degree and duration of overload, and provides clear criteria for the implementation of tiered strategies.

[0013] Preferably, the heavy load state determination adopts a sliding comparison mechanism between the point-by-point power sequence and the available power in the predicted time domain. The power excess amount, excess duration, and excess starting position are used as multi-dimensional criteria to distinguish between short-term power spikes caused by local hard soil blocks on the ground and continuous power overload caused by continuous increase in tillage depth or driving on slopes. This enables intelligent differentiation of overload types, avoids accidental speed reduction affecting normal operation, and refines the classification of heavy load modes, providing accurate basis for differentiated power coordination.

[0014] Preferably, step 5 specifically includes: When the load condition is determined to be normal, the walking power and rotary tillage power predicted by IGWO-TSMixer are directly used as the power command of the underlying motor controller, and a first-order low-pass filter is performed on the total power demand to smooth the battery power output, suppress the impact of high-frequency current fluctuations on battery life, maintain stable power output, and reduce current surges to extend battery cycle life. When a short-term heavy load is detected, priority coefficients for the walking system and rotary tillage system are introduced. Power commands are dynamically allocated between the walking and rotary tillage systems based on a weighted ratio of available battery power. The ratio of the two coefficients is adaptively adjusted according to the relative magnitude of the current traction resistance and PTO torque to adaptively allocate limited power, taking into account both walking stability and rotary tillage quality. When the system is determined to be under continuous heavy load, an additional speed correction is calculated based on the power coordination and allocation. The speed reduction amplitude is obtained by multiplying the ratio of the power excess to the available power by the speed correction coefficient. The target travel speed is reduced in advance to reduce the traction power demand and reduce the instantaneous load impact on the rotary tiller shaft. The active speed reduction reduces the peak power, alleviates battery overload, and ensures system safety.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This method for battery power allocation in rotary tillage based on short-term load prediction of IGWO-TSMixer constructs a TSMixer multivariate time-series prediction model based on a pure MLP architecture. It combines time mixing and feature mixing modules to extract the time-series dependencies and variable interaction relationships in the historical load sequence, enabling direct multi-step prediction of future short-term traction resistance and rotary tillage load. An improved gray wolf optimization algorithm is introduced to adaptively optimize the key hyperparameters of TSMixer, so that the model can maintain stable prediction accuracy under different tillage depths and load fluctuations, significantly enhancing the predictive model's adaptability to variable soil and operating conditions.

[0016] 2. This battery power allocation method for rotary tillage operations based on short-term load prediction of IGWO-TSMixer utilizes the IGWO-TSMixer model to obtain the trend of future short-term load changes in advance, transforming the traditional passive response mode of current load into an active prediction mode of future load. This enables the battery power allocation system to have forward-looking perception capabilities. By knowing the fluctuation trend of traction resistance and rotary tillage load in advance, the controller can adjust the power output strategy in advance, providing a reliable information feedforward basis for optimizing the energy consumption of the whole machine and improving the dynamic response performance.

[0017] 3. The battery power allocation method for rotary tillage based on IGWO-TSMixer short-term load prediction introduces battery SOC safe operating range constraints and a real-time calculation method for available power in power allocation. It dynamically corrects the battery's available power according to the SOC state, avoiding the battery from being subjected to excessively high discharge rates under low charge conditions. At the same time, through a power smoothing filtering strategy under normal load, it effectively suppresses high-frequency current fluctuations, reduces the rate of change of battery output current, and reduces the adverse effects of current surges on battery life. Under the premise of meeting the operating power requirements, it fully ensures the safety of the battery system and extends the service life of the power battery.

[0018] 4. The battery power allocation method for rotary tillage based on short-term load prediction of IGWO-TSMixer classifies the load state into normal load, short-term heavy load and continuous heavy load according to the comparison results of predicted load and battery available power. It then implements a graded power coordination strategy for different states. Under heavy load conditions, power allocation is carried out by dynamically adjusting the priority coefficients of the walking system and the rotary tillage system to make the available power reasonably allocated between the two. Under continuous heavy load, the target travel speed is reduced in advance to actively reduce power demand, effectively resolve power allocation conflicts, prevent vehicle stall and sudden drop in cutter shaft speed, and ensure the stability of the whole machine and the quality of rotary tillage operation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the workflow of the rotary tillage battery power allocation method based on IGWO-TSMixer short-term load prediction of the present invention. Figure 2 This is a schematic diagram of the method flow for the rotary tillage battery power allocation method based on IGWO-TSMixer short-term load prediction according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Example 1, please refer to Figure 1 , Figure 2This invention provides a technical solution: a battery power allocation method for rotary tillage operations based on short-term load prediction using IGWO-TSMixer, the specific steps of which are as follows: Step 1: Collect multi-source time-series data on driving speed, actual tillage depth, and PTO torque within 2-5 seconds of historical data during rotary tillage operations. Construct a sliding time window sample set, and divide it into training, validation, and test sets according to time sequence. Build a standardized sample library covering multiple working conditions to lay the data foundation for subsequent model training and generalization evaluation. Simultaneously collect data on the electric horticulture platform's driving speed, actual tillage depth, platform roll and pitch angles, traction resistance, PTO torque, rotary tillage motor power, battery voltage, battery current, and battery power under rotary tillage conditions at a fixed sampling period. This forms a multi-source time-series data matrix, realizing the coordinated acquisition of multiple physical quantities and providing a complete state information foundation for load prediction. A sliding time window is used to construct a supervised learning model. The sample set is designed with a historical input window length of 2–5 seconds and a future prediction window length of 2–5 seconds. Multi-source data from the historical window corresponding to each sampling time is used as input features, and the traction resistance, PTO torque, and rotary tillage power sequences within the future window are used as prediction labels to generate a sample set. This supports the model in learning temporal causal relationships and improves prediction reliability. The sample set is divided into a 70% training set, a 15% validation set, and a 15% test set in chronological order to ensure that data within the same time window does not cross sets. Training set statistics are calculated separately and used for normalization processing of all subsequent data groups to avoid leakage of time series information, ensure the independence and objectivity of each stage, prevent data from crossing sets and causing inflated performance, and ensure fair evaluation. It should be noted that the sampling period is preferably set to 0.05s to capture the high-frequency load transient characteristics caused by changes in soil firmness and tillage depth fluctuations during rotary tillage, while avoiding data redundancy. The sensor arrangement is as follows: the travel speed is obtained through a Hall speed sensor installed on the drive wheel hub; the actual tillage depth is measured in real time through an ultrasonic ranging sensor installed on the front crossbeam of the platform; the platform attitude is obtained through a six-axis inertial measurement unit; the traction resistance is measured through a tension / compression sensor installed at the traction pin; the PTO torque is obtained through a torque sensor installed at the input end of the rotary tiller shaft; the rotary tiller motor power, battery voltage, battery current, and battery power are all read in real time through the voltage and current sampling modules built into the motor controller and battery management system. All sensor signals are synchronously acquired by the data acquisition card and converted into digital signals. The length of the historical input window and the length of the future prediction window are both preferably set to 2s, corresponding to 40 sampling steps, to balance the effective capture of load change trends and the rapid response requirements of the prediction model. The specific construction method is as follows: at the current moment... The model input is the past Multi-source payload sequence at each sampling point , can be represented as: , The number of input features determines the model output. The traction resistance, rotary tillage PTO torque, and rotary tillage power sequences at each sampling point. The sliding step size is set to a single sampling period, thereby generating a large number of overlapping samples. Each sample consists of a historical input window and a future prediction window. The historical window contains the multi-source load state over the past 2–5 seconds, and the prediction window contains the traction resistance and rotary tillage load response over the next 2, 3, or 5 seconds. Let the first... The input for each sample is The output is Then the sample set can be represented as: , The sample size is as follows: For each tillage depth condition (5cm, 10cm, 15cm), 3600 consecutive sampling points are used. First, the first 70% (sampling points 1-2520) are selected chronologically as the training set, the middle 15% (sampling points 2521-3060) as the validation set, and the last 15% (sampling points 3061-3600) as the test set. The data for each of the three tillage depth conditions are divided independently, without mixing samples from different conditions. The normalization parameter for the training set is calculated as follows: In the formula, and These are the mean and standard deviation of the corresponding variables in the training set, respectively. The validation and test sets must be normalized using the statistics of the training set to avoid leakage of test information. For abnormal jump points of the sensor, the three-standard-deviation criterion or moving median filtering can be used to remove them. For short-term missing data, linear interpolation or forward imputation methods can be used to recover them.

[0022] Step 2: Construct a TSMixer multivariate time-series prediction model based on a pure MLP architecture. Through time mixing and feature mixing modules, extract time-series dependencies and variable interactions to directly predict traction resistance and rotary tillage load for the next 2-5 seconds in multiple steps, obtaining future load trends in advance. TSMixer, the time-series mixer, enables lightweight multi-step prediction, providing real-time and reliable load change trend information for power allocation. The TSMixer multivariate time-series prediction model, based on a pure MLP architecture, simultaneously feeds multi-source load sequences from historical input windows into the time mixing and feature mixing modules. Through fully connected mapping across time steps, it extracts short-term trends and periodic fluctuation dependencies between different historical moments, accurately capturing short-term load fluctuation patterns and providing support for power feedforward control. To provide reliable time-series features, the intermediate feature sequence output by the time mixing module is fed into the feature mixing module. Fully connected mapping is performed along the feature dimension to extract the variable interaction relationship between driving speed, actual tillage depth, traction resistance, PTO torque, and battery state variables. Residual connections and layer normalization are used to enhance information transmission efficiency, effectively integrate multi-variable coupled information, and improve the model's comprehensive perception of complex working conditions. Multiple TSMixer modules are stacked to build a deep prediction network. The final encoded time-series features are mapped through a linear prediction head to the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence for each prediction step within the next 2 to 5 seconds. This enables direct multi-step output of multi-variable time-series data, outputting multi-step prediction results at once, avoiding the accumulation of recursive errors, and improving prediction efficiency and stability. It should be noted that the TSMixer multivariate time series prediction model receives a standardized multi-source time series data matrix with dimensions of 40×10. Here, 40 corresponds to the number of sampling steps within a 2-second historical window, and 10 corresponds to the number of input features. The time mixing module, as the first-level processing unit of the TSMixer multivariate time series prediction model, performs a fully connected mapping along the time dimension, enabling the model to capture the short-term fluctuations in traction resistance and PTO torque during rotary tillage operations as soil conditions change. Specifically, the input matrix is ​​transposed along the time axis and fed into the time mixing fully connected layer. This fully connected layer independently performs a linear transformation across 40 time steps for each feature dimension, establishing global temporal dependencies between historical moments. The feature mixing module receives the intermediate feature sequence output by the time mixing module, performs a fully connected mapping along the feature dimension, and independently performs a linear transformation across the feature dimension for each time step, mapping the 10-dimensional input features to hidden dimension features. This allows the model to integrate the coordinated changes in speed, resistance, torque, and power. Residual connections directly superimpose the module inputs and feature mixture outputs, while layer normalization standardizes the feature distribution. Both work together to alleviate gradient vanishing and accelerate model convergence. The TSMixer deep prediction network is constructed using a module stacking approach, with a preferred stacking layer of 1 to maintain model lightweightness and meet embedded deployment requirements. Each TSMixer module sequentially performs time mixing and feature mixing operations, with the output feature sequence of the previous module serving as the input sequence of the next module. Through progressive information extraction, the model gradually constructs a hierarchical representation from the original load sequence to high-level abstract features, ensuring that the final encoded features simultaneously contain short-term fluctuation details and long-term trend information. The linear prediction head uses a fully connected layer structure to directly map the final encoded features to a three-dimensional output sequence for the next 40 sampling steps (corresponding to 2 seconds), including 40 predicted values ​​each for traction resistance, PTO torque, and rotary tillage power. Furthermore, step 2 also includes: The multi-source time-series data matrix input to the current TSMixer module is normalized by layers and then transposed along the time dimension. After transposition, the time step dimension of the matrix is ​​transformed into the feature dimension, and the feature dimension is transformed into the time step dimension. The transformation result is input into a time-mixing fully connected layer for fully connected mapping between time steps. The mapping result is reversed and restored, then superimposed with the original multi-source time-series data matrix via residual connection to obtain a time-mixed output. This enhances the efficiency of time-series feature transfer and improves the modeling accuracy of dependencies between historical moments. The time-mixed output is normalized by layers and then input along the feature dimension into a feature-mixing fully connected layer for fully connected mapping between variables, extracting the transitions between driving speed, actual tillage depth, traction resistance, PTO torque, and battery state variables. The coupling relationship between quantities is mapped, and the result is superimposed through residual connection to obtain the feature mixture output, which completes the information integration of a single TSMixer module, strengthens the collaborative expression of multiple variables, and improves the model's overall perception of complex operating conditions. Multiple TSMixer modules are stacked sequentially, and the feature sequence output by the previous module is used as the input sequence of the next module. Through the alternating mapping of progressive time mixing and feature mixing, a multi-level abstract representation of the historical load sequence from short-term fluctuations to long-term trends is constructed to obtain future load trends in advance. Finally, the fully connected prediction head directly maps to the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence of each prediction step in the next 2 to 5 seconds, abstracting the time series features layer by layer, and enhancing the model's comprehensive expression of the load evolution law. It should be noted that the current TSMixer module input matrix stored in the data buffer is passed to the layer normalization unit for feature distribution standardization. The matrix is ​​then transposed along the time axis by the transpose register, enabling the fully connected layer to perform linear mapping along the time dimension. The transposed matrix is ​​fed into the temporal mixing fully connected computation array, which contains a 40×40 weight matrix. Weighted summation is performed independently across time steps for each feature channel, establishing global dependencies between different moments in the historical load sequence. After the mapping result is reversed and restored, it is element-wise superimposed with the original input matrix through a residual accumulator to form a temporal mixing output feature map, while retaining both the original input information and the temporal features extracted along the time dimension. The temporal mixing output feature map enters the second layer normalization unit for distribution standardization. The standardized feature map is then fed along the feature dimension into the feature mixing fully connected layer, which uses a 10×10 weight matrix for each time step. Each step independently performs a linear transformation across feature dimensions, mapping the 10-dimensional input features to hidden dimension features. It comprehensively extracts the synergistic change patterns between speed, drag, torque, and power. The feature mixing mapping result is superimposed with the temporal mixing output through residual connections. The fused feature map serves as the final output of the current TSMixer module, completing the information integration of a single module. Each TSMixer module is called sequentially according to the preset module layer number. The feature sequence output by the first module is used as the input data for the second module, and so on, completing the layer-by-layer transmission. Each module alternately performs temporal mixing and feature mixing operations, enabling the network to gradually build a hierarchical representation from the original load details to high-level abstract features. Each layer focuses on extracting the change patterns at different time scales. After all modules have been calculated, the encoded features output by the last layer are flattened and sent to the fully connected prediction head. The prediction head directly converts the feature vector into the three-dimensional output sequence of each future prediction step through a single-layer linear mapping. Furthermore, the TSMixer multivariate time series prediction model adopts a pure MLP architecture to replace the recursive and attention structure. Its time mixing and feature mixing dual branches are both based on fully connected layers to achieve forward mapping, which simplifies the calculation process, reduces the deployment difficulty, and improves the real-time prediction response speed. Moreover, the time mixing and feature mixing dual branches perform cross-parameter full connection along the variable dimension, directly modeling the instantaneous coupling relationship between speed-resistance-torque-power, strengthening the ability to express the instantaneous correlation between physical variables, and improving the consistency of multi-objective prediction.

[0023] Step 3: Introduce the Improved Gray Wolf Optimization Algorithm (IGWO). Encode the learning rate, Dropout coefficient, and hidden layer dimension of TSMixer as gray wolf positions. Use the root mean square error of the validation set as the fitness to adaptively optimize hyperparameters. The improved gray wolf optimization algorithm is IGWO, which enhances the generalization stability under different tillage depths, reduces the cost of manual parameter tuning, and improves the consistency of predictions across different conditions. Encode the initial learning rate, L2 regularization coefficient, Dropout coefficient, number of module stacking layers, temporal mixture hidden dimension, and feature mixture hidden dimension of the TSMixer multivariate time series prediction model as the position vector of individual gray wolves in the improved gray wolf optimization algorithm. Set the search boundary range of each hyperparameter, and encode the hyperparameter combination as a position vector to ensure that the search boundary is reasonable and realize the systematic search of key hyperparameters, avoiding... To avoid the blindness of manual trial and error, the fitness function is the weighted average of the three errors of traction resistance prediction RMSE, PTO torque prediction RMSE and rotary tillage power prediction RMSE on the validation set. A cosine adaptive convergence factor is used to dynamically adjust the intensity of global exploration and local development. A particle swarm velocity memory mechanism is introduced to retain historical search directions, dynamically balance global exploration and local development, improve search efficiency, retain historical search directions, accelerate convergence and improve stability. A random difference mutation strategy is applied to the current optimal gray wolf position to generate new candidate solutions. Individuals with better fitness are retained through greedy selection. The gray wolf population is iteratively updated until the maximum number of generations is reached. The optimal hyperparameter combination that minimizes the multi-task prediction error on the validation set is output, which enhances population diversity, suppresses premature convergence, outputs the global optimal solution and improves the generalization ability of cross-working condition prediction. It should be noted that the six key hyperparameters of the TSMixer multivariate time series prediction model are arranged sequentially to form the position vector of the individual gray wolf. The search space of each hyperparameter is set independently based on prior knowledge and model sensitivity analysis. The initial learning rate search interval is set to [0.0001, 0.01], and logarithmic uniform sampling is used to cover different orders of magnitude; the L2 regularization coefficient search interval is set to [1e-6, 1e-2] to control model complexity and prevent overfitting; the Dropout coefficient search interval is set to [0.05, 0.5] to randomly drop neurons to enhance generalization ability; and the module stacking layer... The search term is within the integer range {1, 2, 3} to balance model depth and computational cost. The temporal mixture hidden dimension is discretely selected from {32, 64, 80, 128}, and the feature mixture hidden dimension is discretely selected from {64, 128, 256, 512}. Boundary constraints on each hyperparameter ensure that all candidate solutions are within the feasible region during the search process. The fitness function is defined as a weighted sum of three RMSEs. The weight coefficients are set to 0.3, 0.3, and 0.4 respectively, based on the dimensions of each physical quantity and actual control requirements, to highlight the importance of rotary tillage power in energy management. The expression for the cosine adaptive convergence factor is: ,in This represents the current iteration number. To maximize the number of iterations, this factor approaches 2 in the early stages of iteration to enhance global exploration capabilities, and approaches 0 in the later stages to promote fine-grained local searches. The particle swarm velocity memory mechanism introduces an inertial weight w=0.7, allowing individual gray wolves to retain their velocity components from the previous moment when updating their positions, thus preventing abrupt changes in search direction. The specific operation of the random difference mutation strategy is as follows: two distinct gray wolf individuals are randomly selected from the current population; their position vectors are subtracted and multiplied by a mutation scaling factor F (F randomly selected within the interval [0.2, 0.8]), and then compared with the current best gray wolf. Position vectors are superimposed to generate candidate solutions for mutation. A greedy selection mechanism compares the fitness value of the candidate solution with that of the original optimal solution. If the mutated solution has a lower fitness function value, it replaces the original optimal solution; otherwise, the original solution is retained. The population size is set to 20, and the maximum number of iterations is set to 50. Under these settings, the optimization process usually converges within 30 generations. The entire optimization process is executed independently on the validation set to ensure that the hyperparameter optimization process does not touch the test set data. Finally, the output is a set of hyperparameter combinations that minimizes the multi-task prediction error on the validation set.

[0024] Step 4: Utilize the optimized IGWO-TSMixer to output the predicted future load value, combine the traction resistance and PTO torque to estimate the future power demand of the walking and rotary tillage systems, and calculate the current available power based on the battery SOC and current limit to determine whether the system has entered a heavy load state. This will help identify power supply and demand contradictions in advance and provide a reliable triggering basis for active coordinated control under heavy load conditions.

[0025] Step 5: Based on the judgment results, implement a graded power coordination strategy: smooth the output under normal load, dynamically adjust the priority coefficients of walking and rotary tillage under continuous heavy load, and reduce the target driving speed in advance to achieve stable battery power distribution and work quality optimization, realize the coordinated control of power smoothness and work stability, and take into account both battery safety and rotary tillage work quality.

[0026] Example 2, as Figure 1 , Figure 2As shown, based on Embodiment 1, this invention provides a technical solution: Step 4 specifically includes: inputting real-time collected historical window multi-source data into the optimized IGWO-TSMixer prediction model. The model directly outputs the predicted values ​​of traction resistance, PTO torque, and rotary tillage power for each prediction step within the next 2-5 seconds, serving as a feedforward basis for forward-looking power allocation. This enables the advanced acquisition of load trends, provides time margin for power allocation, reduces the risk of response lag, and enhances active control capabilities. The future driving power of the walking system is estimated based on the predicted traction resistance and current driving speed, and the PTO torque is calculated based on the predicted PTO torque. The predicted torque value and the angular velocity of the cutter shaft are used to estimate the future operating power of the rotary tillage system. The sum of the two values ​​and the power of the auxiliary system are added to obtain the future total battery power demand sequence. The power demand is quantified in advance, providing clear indicators for battery scheduling and improving the foresight and accuracy of energy management. The current battery SOC, terminal voltage and bus current are collected. The current available power of the battery is calculated based on the SOC correction factor and the maximum allowable discharge current. The future total power demand sequence is compared with the available power step by step to determine whether each prediction step is under heavy load. The power margin is evaluated in real time, and the overload risk is accurately identified, providing a reliable triggering basis for graded power coordination. It should be noted that, in one specific embodiment of the present invention, the standardized historical data vector (dimension 40×10) collected in real time is input into the IGWO-TSMixer prediction model that has been trained offline. The forward computation of the model is performed by the inference engine deployed in the ARM embedded controller. Each TSMixer module completes the layer-by-layer operation according to the defined temporal mixing-feature mixing order. Finally, the linear prediction head outputs the traction resistance, PTO torque and rotary tillage power prediction vectors for the next 40 sampling steps (corresponding to 2 seconds) in parallel. The output delay does not exceed one sampling period (0.05 seconds), which meets the real-time requirements of forward control of rotary tillage operation. The predicted values ​​of traction resistance and PTO torque for the next 40 steps are read sequentially from the prediction buffer, and combined with the travel speed and cutter shaft angular velocity at the current sampling time, they are divided by the efficiency of the travel drive system ( ) and rotary tillage transmission efficiency ( After obtaining point-by-point estimates of the driving power and rotary tillage power, these are summed and then added to the auxiliary system power (including the steering motor, cooling fan, and controller standby power consumption, taken as 150W) to form a 40-step sequence of future total power demand. The power value at each step is stored as a floating-point number in the power prediction buffer. The SOC percentage, terminal voltage, and bus current broadcast by the battery management system are read via the CAN bus at a synchronization cycle of 0.05 seconds. Based on the current SOC value, the SOC correction coefficient kSOC is obtained from a table (kSOC = 1.0 when SOC is higher than 30%). The SOC decreases linearly to 0.7 in the 20% to 30% range, and is limited to 0.5 when it is below 20%. This value is multiplied by the terminal voltage and the maximum allowable discharge current (100A) specified in the battery datasheet to obtain the current available power value. Then, the total power demand sequence for the next 40 steps is subtracted from the current available power value one by one. If there are more than 3 consecutive steps where the power demand exceeds the available power and the excess is greater than 50W, the prediction window is marked as an overload state, and the first excess position and the maximum excess are recorded. A three-element judgment vector containing the excess flag, the number of consecutive excess steps, and the maximum power deficit is generated. Different forecast scales correspond to different control effects, as shown in Table 1 below; Table 1. Different prediction scales and their control effects Therefore, the model can adopt a multi-step prediction method and output load prediction results for 2s, 3s and 5s at the same time. It can make hierarchical decisions based on the prediction results of different time scales: short time scale is used for rapid battery protection, medium time scale is used for power redistribution, and long time scale is used for coordinated optimization of operation speed and operation power. Furthermore, step 4 also includes: setting the battery SOC safe operating range and maximum allowable discharge current; the battery's available power is calculated by multiplying the current terminal voltage, the maximum allowable current, and the SOC correction factor; when the SOC is below the lower threshold, the correction factor is linearly reduced to limit the discharge rate; the upper limit of available power is dynamically adjusted to prevent over-discharge at low SOC from damaging battery life; when estimating the power of the walking drive system, the efficiency of the walking drive system is considered, and it is obtained by dividing the product of the predicted traction resistance value and the predicted driving speed value by the efficiency; when estimating the power of the rotary tillage operation, the transmission efficiency of the rotary tillage operation is considered, and it is obtained by dividing the product of the predicted PTO torque value and the predicted cutter shaft angular velocity value by the efficiency; a transmission loss correction is introduced to improve the actual consistency and reliability of the power estimation; when the total power demand in any prediction step in the future prediction time domain is greater than the battery's available power, the prediction window is marked as a heavy load state, and the heavy load duration and the maximum power excess are recorded, providing a basis for graded decision-making for subsequent power coordination strategies, accurately quantifying the degree and duration of overload, and providing clear criteria for the implementation of graded strategies; It should be noted that the safe operating range of the battery's State of Charge (SOC) is preset to 20%–100%, and the maximum allowable discharge current is set to 100A according to the battery specifications. The BMS periodically broadcasts the SOC and terminal voltage via the CAN bus. The controller reads the above data during each sampling interrupt and calculates the SOC correction coefficient in real time. This SOC correction coefficient remains at 1.0 when the SOC is above 30%, decreases linearly in the 20%–30% range, and is limited to 0.5 when the SOC is below 20%. This SOC correction coefficient directly participates in the multiplication calculation of the battery's available power, causing the available power to decrease as the SOC decreases, thereby achieving the control objectives of limiting the discharge rate under low charge conditions, protecting battery life, and preventing voltage drops; the efficiency of the walking drive system… The transmission efficiency for rotary tillage is set at 0.85. The value is set to 0.82. Both values ​​are obtained based on bench calibration tests. The controller reads the predicted values ​​of traction resistance and PTO torque for the next 40 steps sequentially from the prediction buffer. At the same time, it obtains the travel speed and cutter shaft angular velocity from the current sensor readings. Assuming that the speed and angular velocity remain unchanged within the 2-second prediction time domain, the controller performs division operations to obtain the point-by-point estimation sequence of travel power and rotary tillage power. The total power demand sequence for the 40 steps is compared with the current available battery power one by one. If there are more than 3 consecutive steps where the power demand exceeds the available power and the excess is greater than 50W, the overload flag is set, and the first overload position, the number of consecutive overload steps, and the maximum power deficit are recorded. This judgment logic is executed once every 0.05 seconds. The generated ternary judgment vector is passed for power coordination and allocation to distinguish between short-term power spikes caused by local hard soil blocks and continuous power overload caused by continuous increase in tillage depth or driving on slopes. Based on the traction resistance prediction results, the future driving power can be expressed as: ; In the formula, For the future Predicted walking drive power value at each sampling point For the future Predicted traction resistance values ​​at each sampling point For the future The driving speed at each sampling point For the efficiency of the walking drive system; Based on the predicted PTO torque of rotary tillage, the future rotary tillage power can be expressed as: ; In the formula, For the future Predicted power values ​​for rotary tillage operations at each sampling point For the future Predicted PTO torque values ​​for rotary tillage at each sampling point For the future The rotary tiller shaft angular velocity at each sampling point To improve the transmission efficiency of rotary tillage operations; If the rotary tiller motor power has been directly collected, the predicted rotary tiller power output by TSMixer can also be used directly as the rotary tiller power requirement. Therefore, the total future battery power requirement is: ; In the formula, For the future The predicted total battery power demand at each sampling point For auxiliary system power; Further define the power ratio of the walking system and the rotary tillage system in the future prediction time domain: ; ; in, For the future The percentage of walking system power at each sampling point For the future The power percentage of the rotary tillage system at each sampling point, when When the load continues to rise, it indicates that the rotary tiller shaft load has become the main source of power demand; when A significant increase indicates that factors such as traction resistance, gradient, or track slippage have a greater impact on the chassis's travel power demand. In addition, the heavy load condition determination adopts a sliding comparison mechanism between the point-by-point power sequence and the available power in the predicted time domain. The power excess amount, excess duration and excess starting position are used as multi-dimensional criteria to distinguish between short-term power spikes caused by local hard soil blocks on the ground and continuous power overload caused by continuous increase in tillage depth or driving on slopes. This enables intelligent differentiation of overload types, avoids accidental speed reduction affecting normal operation, and refines the classification of heavy load modes, providing accurate basis for differentiated power coordination.

[0027] Step 5 specifically includes: When the load is determined to be normal, the walking power and rotary tillage power predicted by IGWO-TSMixer are directly used as the power command of the underlying motor controller, and a first-order low-pass filter is performed on the total power demand to smooth the battery power output, suppress the impact of high-frequency current fluctuations on battery life, maintain stable power output, and reduce current surges to extend battery cycle life. When the load is determined to be short-term heavy load, the priority coefficients of the walking system and the rotary tillage system are introduced, and the power command is dynamically allocated between the walking and rotary tillage systems according to the weighted ratio of available battery power. The ratio of the two coefficients is adaptively adjusted according to the relative magnitude of the current traction resistance and PTO torque to adaptively allocate limited power, taking into account both walking stability and rotary tillage quality. When the load is determined to be continuous heavy load, the speed correction amount is calculated on the basis of power coordination allocation. The speed reduction amplitude is obtained by multiplying the ratio of the power excess to the available power by the speed correction coefficient. The target driving speed is reduced in advance to reduce the traction power demand and reduce the instantaneous load impact of the rotary tillage cutter shaft. The speed reduction actively reduces the peak power, alleviates battery overload, and ensures system safety. When the total power demand is predicted to exceed the available battery power within the next 2–5 seconds: ; In the formula, The projected total future battery power demand is the sum of driving power, rotary tillage power, and auxiliary power. This represents the battery's current available power, which is constrained by factors such as SOC, voltage, and current. Power coordination and allocation between the walking system and the rotary tillage system are required. Considering the quality of rotary tillage operations and the stability of platform travel, a priority coefficient method is adopted: ; ; In the formula, This is the final power command for the walking system, i.e., the power actually allocated to the walking motor. This is the final power command for the rotary tiller system, i.e., the power actually allocated to the rotary tiller motor. Predict the power demand for the walking system. To predict the power demand of the rotary tillage system, This represents the priority coefficient for the walking system. This is the priority coefficient for the rotary tillage system. When prioritizing the quality of rotary tillage operations, it can be set as follows: Prioritizing the power output of the rotary tiller shaft to prevent a decrease in shaft speed leading to poorer soil pulverization quality, the system will increase power output when a sustained increase in traction resistance, a higher risk of track slippage, or a significant decrease in vehicle speed is anticipated. Prioritize ensuring the traction capability of the walking system; It should be noted that power distribution under normal load conditions is automatically executed by the power coordination controller. It reads the power sequence for the next 40 steps of walking and rotary tillage from the prediction buffer and directly uses it as the feedforward command for the motor controller. Simultaneously, a first-order low-pass filter smooths the total power demand, with a filtering time constant set to 0.2 seconds. This filters out high-frequency power fluctuations caused by localized soil hardening or minor terrain changes, keeping the battery output current change rate below 5A / s and ensuring the motor speed response meets operational requirements. Short-term heavy load conditions are handled by the priority arbitration module in the power coordination controller. Based on the real-time sampled ratio of traction resistance to PTO torque, the priority coefficient of the walking system is dynamically adjusted. Priority coefficient of rotary tillage system The ratio of the two is continuously adjustable within the range of 0.3 to 0.7. When the traction resistance suddenly increases to more than 1.5 times the PTO torque, the controller automatically... Increase it to 0.6 or higher to prioritize driving force and prevent vehicle stalling; when PTO torque is dominant, then... The power distribution is adjusted to above 0.6 to ensure the quality of soil crushing, and the adaptive adjustment response time is less than two sampling periods to ensure that the power distribution can track load changes in a timely manner. Under continuous heavy load conditions, the speed look-ahead correction module is triggered to intervene in the control. This module takes the relative ratio of power deficit as input and multiplies it by the speed correction coefficient. The forward velocity correction amount was calculated. and the corrected target speed Send to the walking drive controller, The original target travel speed is maintained, while the rotary tillage system power command remains unchanged to ensure stable cutter shaft speed and guarantee work quality. Speed ​​correction is updated every 0.05 seconds, with the speed reduction limited to within 0.1 m / s to prevent sudden speed changes that could cause platform instability. Once the power deficit disappears, the speed is adjusted to 0.02 m / s. 2 The slope ratio was restored to the original target value; By reducing the driving speed, the traction power demand in the future prediction time domain can be reduced, keeping the battery output within a safe range, while reducing the instantaneous load impact on the rotary tiller shaft. The battery power distribution status and control rules are shown in Table 2 below. Table 2 Battery Power Distribution Status and Control Rules in, This represents the battery's current state of charge, indicating the percentage of its rated capacity that remains usable. This is the minimum permissible state of charge threshold for a battery, representing the lower limit for safe battery operation.

[0028] Example 3, as Figure 1 , Figure 2 As shown, based on Examples 1 and 2, the evaluation metrics for the prediction model include root mean square error, mean absolute error, and coefficient of determination: ; ; ; in, These are measured values. For predicted values, The average value is the measured value. In addition to the overall prediction accuracy, the impact of the prediction results in the next 2s, 3s and 5s on the power allocation strategy should be evaluated. In addition, the evaluation indicators are also developed from three aspects: battery power stability, operation stability and energy consumption, as shown in Table 3 below, which is the evaluation index of predictive control and battery power distribution effect. Table 3 Evaluation Indicators for Predictive Control and Battery Power Distribution The rate of reduction in battery current fluctuation can be defined as: ; In the formula, For the reduction rate of battery current fluctuation, This refers to the battery current under no-prediction strategy. To predict battery current under the power allocation strategy based on TSMixer, The standard deviation of battery current under no prediction strategy. The standard deviation of battery current based on the TSMixer predicted power allocation strategy; The work power satisfaction rate can be defined as: ; In the formula, The power satisfaction rate for rotary tillage operations is used to evaluate the extent to which the power actually allocated to the rotary tillage system can meet its needs under conditions of limited battery power. The closer this value is to 100%, the better the quality of operation. For the first The actual commanded power of the rotary tillage system at each sampling point For the first Power requirement of the rotary tillage system at each sampling point; This indicator is used to evaluate whether the predictive control strategy can ensure the power requirements of rotary tillage while protecting the battery; To verify the effectiveness of the proposed method, the load prediction part sets up four sets of algorithms for comparison: TSMixer, IGWO-TSMixer, BiLSTM-Transformer, and IGWO-BiLSTM-Transformer. The four models use completely consistent data partitioning, input and output variables, prediction scale, and evaluation index. The focus is on comparing the impact of different time series modeling structures and IGWO hyperparameter optimization on the generalization performance of the test set. The power allocation part continues to set up three strategies: no prediction power following, ordinary TSMixer prediction allocation, and IGWO-TSMixer look-ahead allocation. The evaluation is comprehensively evaluated by indicators such as prediction error, battery current fluctuation, energy consumption per unit operating area, driving speed fluctuation, and PTO speed fluctuation. The algorithm was validated using measured data from rotary tillage operations. The data included three target tillage depths of 5cm, 10cm, and 15cm, with 3600 sampling points for each condition and a sampling interval of 0.05s. Within each condition, the training set, validation set, and test set were divided into three groups according to time sequence at a ratio of 70%, 15%, and 15%, respectively. Sliding windows were constructed to prevent information leakage caused by windows crossing different conditions or random shuffling. The model inputs included driving speed, actual tillage depth, traction resistance, PTO torque, rotary tillage power, battery voltage, battery current, SOC, and total power demand. Three types of loads were predicted for the next 40 points using 40 historical data points. The test set performance comparison of the four short-term load prediction models is shown in Table 4 below. Table 4. Performance Comparison of Four Short-Term Load Prediction Models on the Test Set Table 4 shows that the IGWO-TSMixer achieves the best overall performance, with RMSE values ​​of 101.33 N for traction resistance, 9.93 N·m for PTO torque, and 266.03 W for rotary tillage power, and an average RMSE of 125.77. Compared to the ordinary TSMixer, the average RMSE is reduced by 2.29%. The BiLSTM-Transformer has an average RMSE of 158.10, with RMSE values ​​of 13.26 N·m for PTO torque and 352.51 W for rotary tillage power, indicating that the recurrent encoding and attention structure failed to surpass the lightweight TSMixer under the current sample size and 40-step direct multi-output conditions. The validation set fitness of the IGWO-BiLSTM-Transformer is improved, but the average RMSE on the test set increases to 164.45, and the traction resistance R... 2 The score dropped to 0.640, indicating that the optimization results overfitted to the validation period and have not yet formed a stable cross-period generalization advantage. In addition, the performance improvement of IGWO-TSMixer relative to TSMixer is shown in Table 5 below. Table 5 Performance improvement of IGWO-TSMixer relative to TSMixer Further comparison reveals that IGWO's effects on different base networks are inconsistent. TSMixer has fewer parameters and a clear time and feature mixing path, allowing IGWO to achieve a small but stable error reduction with a limited search budget. BiLSTM-Transformer, which includes recurrent states, attention weights, and a prediction head, has stronger parameter coupling, and its shorter fitness training tends to favor specific validation segments. Therefore, IGWO-TSMixer, which has the best overall performance on the test set, is used as the load predictor for battery power look-ahead allocation. For IGWO-BiLSTM-Transformer, it is advisable to increase the wolf pack size and number of iterations in the future, and use rolling cross-validation, repeated random seeds, and model complexity penalties to jointly constrain fitness, so as to reduce the risk of search overfitting. The validation results of battery power allocation based on predicted load are shown in Table 6 below. Table 6. Verification results of battery power allocation based on predicted load In the battery power distribution verification, based on the actual measured SOC of the prototype of about 58.8%, the battery voltage of about 62.8V and the current limit, the available battery power is about 3233.33W. As shown in Table 6, the first test window did not trigger heavy load power coordination distribution at the prediction scales of 2s, 3s and 5s, and the average power satisfaction rate was 100%. This indicates that the predicted total power demand for walking-rotary tillage in the current test segment is lower than the available battery power. The battery can complete the power supply according to the predicted demand, and a relatively smooth battery reference power is generated through low-pass filtering and power change rate constraint. In summary, the four-model validation results show that the IGWO-TSMixer has the lowest overall error and the highest overall coefficient of determination at the 2s short-term prediction scale, making it more suitable as a feedforward information source for rapid battery protection and short-term power tracking. The BiLSTM-Transformer can describe long-range dependencies, but its computational complexity does not match its generalization benefits under the current data and direct multi-step prediction settings. Its IGWO version also exhibits a validation set selection bias. The 3s and 5s TSMixer results can continue to be used for power redistribution and speed-operational power coordination optimization. Further validation of the heavy-load power distribution effect should be conducted under low SOC, current-limited, or high-load conditions.

[0029] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A battery power allocation method for rotary tillage operations based on short-term load prediction using IGWO-TSMixer, characterized in that, The specific steps are as follows: Step 1: Collect multi-source time-series data of driving speed, actual tillage depth and PTO torque within 2-5 seconds of historical data during rotary tillage operations, construct a sliding time window sample set, and divide it into training set, validation set and test set according to time order; Step 2: Construct a TSMixer multivariate time series prediction model based on a pure MLP architecture to directly predict the traction resistance and rotary tillage load in the next 2 to 5 seconds in multiple steps, and obtain the future load trend in advance. TSMixer is a time series mixer. Step 3: Introduce an improved gray wolf optimization algorithm, which encodes the learning rate, Dropout coefficients, and hidden layer dimensions of TSMixer as gray wolf positions, and uses the root mean square error of the validation set multi-task as the fitness to adaptively optimize the hyperparameters. The improved gray wolf optimization algorithm is IGWO. Step 4: Use the optimized IGWO-TSMixer to output the predicted future load value, combine the traction resistance and PTO torque to estimate the future power demand of the walking and rotary tillage systems, and calculate the current available power based on the battery SOC and current limit to determine whether to enter the heavy load state. Step 5: Based on the judgment results, implement a graded power coordination strategy: smooth the output under normal load, dynamically adjust the priority coefficients of walking and rotary tillage under continuous heavy load, and reduce the target driving speed in advance.

2. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 1, characterized in that: Step 1 specifically includes: The electric gardening platform was synchronously collected at a fixed sampling period, including its driving speed, actual tillage depth, platform roll and pitch angles, traction resistance, PTO torque, rotary tillage motor power, battery voltage, battery current, and battery power under rotary tillage conditions, forming a multi-source time-series data matrix. A sliding time window is used to construct supervised learning samples. The length of the historical input window is set to 2-5s, and the length of the future prediction window is set to 2-5s. The multi-source data of the historical window corresponding to each sampling time is used as the input features, and the traction resistance, PTO torque and rotary tillage power sequence in the future window are used as prediction labels to generate a sample set. The sample set was divided into 70% training set, 15% validation set, and 15% test set in chronological order. The training set statistics were calculated separately and used for normalization of all subsequent data groups.

3. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 1, characterized in that: Step 2 specifically includes: A TSMixer multivariate time series prediction model based on a pure MLP architecture is constructed. The multi-source load sequences of the historical input window are simultaneously fed into the time mixing module and the feature mixing module. The short-term trend and periodic fluctuation dependencies between different historical moments are extracted through a fully connected mapping across time steps. The intermediate feature sequence output by the time mixing module is fed into the feature mixing module, and a fully connected mapping is performed along the feature dimension to extract the variable interaction relationship between driving speed, actual tillage depth, traction resistance, PTO torque and battery state variables. Multiple TSMixer modules are stacked to construct a deep prediction network. The final encoded time-series features are then mapped via a linear prediction head to the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence for each prediction step within the next 2–5 seconds.

4. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 3, characterized in that: Step 2 also includes: The multi-source time series data matrix input to the current TSMixer module is normalized by the layer and then transposed along the time dimension. After transposition, the time step dimension of the matrix is ​​transformed into the feature dimension, and the feature dimension is transformed into the time step dimension. The transformation result is input into the time mixing fully connected layer for full connection mapping between time steps. The mapping result is reversed and restored, and then superimposed with the original multi-source time series data matrix through residual connection to obtain the time mixing output. After the time-mixed output is normalized by the layer, it is input into the feature mixing fully connected layer along the feature dimension to perform fully connected mapping between variables. This extracts the cross-variable coupling relationship between driving speed, actual tillage depth, traction resistance, PTO torque and battery state variables. The mapping results are superimposed by residual connections to obtain the feature mixing output, thus completing the information integration of a single TSMixer module. Multiple TSMixer modules are stacked sequentially, and the feature sequence output by the previous module is used as the input sequence of the next module. Through progressive time mixing and feature mixing, a multi-level abstract representation of the historical load sequence from short-term fluctuations to long-term trends is constructed to obtain future load trends in advance. Finally, the fully connected prediction head directly maps the traction resistance sequence, PTO torque sequence, and rotary tillage power sequence for each prediction step in the next 2 to 5 seconds.

5. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 4, characterized in that: The TSMixer multivariate time series prediction model adopts a pure MLP architecture to replace the recursive and attention structure. Its time mixing and feature mixing dual branches are both based on fully connected layers to achieve positive mapping. Moreover, the time mixing and feature mixing dual branches perform cross-parameter full connection along the variable dimension to directly model the instantaneous coupling relationship between speed, resistance, torque and power.

6. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 1, characterized in that: Step 3 specifically includes: The initial learning rate, L2 regularization coefficient, Dropout coefficient, number of module stacking layers, temporal mixture hidden dimension, and feature mixture hidden dimension of the TSMixer multivariate temporal prediction model are encoded into the position vector of individual gray wolves in the improved gray wolf optimization algorithm, and the search boundary range of each hyperparameter is set. The weighted average of the three errors of traction resistance prediction RMSE, PTO torque prediction RMSE and rotary tillage power prediction RMSE on the validation set is used as the fitness function. The cosine adaptive convergence factor is used to dynamically adjust the intensity of global exploration and local development, and a particle swarm velocity memory mechanism is introduced to retain the historical search direction. Apply a random difference mutation strategy to the current optimal gray wolf position to generate new candidate solutions. Then, retain individuals with better fitness through greedy selection. Iterate and update the gray wolf population until the maximum number of generations is reached. Output the optimal combination of hyperparameters that minimizes the multi-task prediction error on the validation set.

7. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 1, characterized in that: Step 4 specifically includes: The real-time collected historical multi-source data is input into the optimized IGWO-TSMixer prediction model, and the model directly outputs the predicted values ​​of traction resistance, PTO torque and rotary tillage power for each prediction step in the next 2 to 5 seconds. The future driving power of the walking system is estimated based on the predicted traction resistance and the current driving speed. The future working power of the rotary tillage system is estimated based on the predicted PTO torque and the cutter shaft angular velocity. The sum of the two and the power of the auxiliary system are added together to obtain the future total battery power demand sequence. Collect the current battery SOC, terminal voltage, and bus current. Calculate the current available power of the battery based on the SOC correction factor and the maximum allowable discharge current. Gradually compare the future total power demand sequence with the available power to determine whether each prediction step is under heavy load.

8. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 7, characterized in that: Step 4 also includes: Set the battery's safe operating range for SOC and the maximum allowable discharge current. The battery's usable power is calculated as the product of the current terminal voltage, the maximum allowable current, and the SOC correction factor. When the SOC is below the lower limit threshold, the correction factor is linearly reduced to limit the discharge rate. When estimating the power of the walking drive system, the efficiency of the walking drive system is taken into account. The power is obtained by dividing the product of the predicted traction resistance and the predicted travel speed by the efficiency. When estimating the power of the rotary tillage operation, the transmission efficiency of the rotary tillage operation is taken into account. The power is obtained by dividing the product of the predicted PTO torque and the predicted cutter shaft angular velocity by the efficiency. When the total power demand in any prediction step within the future prediction time domain exceeds the available battery power, the prediction window is marked as an overload state, and the overload duration and the maximum power excess are recorded.

9. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 8, characterized in that: The heavy load condition determination adopts a sliding comparison mechanism between the point-by-point power sequence in the predicted time domain and the available power, and uses the power excess amount, excess duration and excess starting position as multi-dimensional criteria to distinguish between short-term power spikes caused by local hard soil blocks on the ground surface and continuous power overload caused by continuous increase in tillage depth or driving on slopes.

10. The method for battery power allocation in rotary tillage operations based on short-term load prediction using IGWO-TSMixer as described in claim 1, characterized in that: Step 5 specifically includes: When the load condition is determined to be normal, the walking power and rotary tillage power predicted by IGWO-TSMixer are directly used as the power command of the underlying motor controller, and a first-order low-pass filter is performed on the total power demand to smooth the battery power output and suppress the impact of high-frequency current fluctuations on battery life. When a short-term heavy load is determined, a priority coefficient for the walking system and a priority coefficient for the rotary tillage system are introduced. Power commands are dynamically allocated between the walking and rotary tillage systems according to a weighted ratio of available battery power. The ratio of the two coefficients is adaptively adjusted according to the relative magnitude of the current traction resistance and PTO torque. When the system is determined to be under continuous heavy load, an additional speed correction is calculated based on the power coordination and allocation. The speed reduction is obtained by multiplying the ratio of the power excess to the available power by the speed correction coefficient. The target travel speed is reduced in advance to reduce the traction power demand and alleviate the instantaneous load impact on the rotary tiller shaft.

Images