An on-line power prediction method for electric tractor based on working condition separation degree
By employing working condition separation characterization and adaptive prediction strategies, the problem of insufficient prediction accuracy and robustness of electric tractors under different operating conditions is solved, achieving efficient online power prediction and improving energy utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI AGRICULTURAL UNIVERSITY
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-21
AI Technical Summary
Existing power prediction methods for electric tractors struggle to balance prediction accuracy and robustness in both steady-state and switching regions under different operating conditions, resulting in low energy efficiency.
By continuously quantifying the separation of operating conditions, an adaptive switching strategy of single or multi-model fusion is adopted, combined with PCA dimensionality reduction and K-Means clustering, to accurately distinguish between steady-state and switching regions and adaptively adjust the prediction model.
It significantly reduces prediction errors, improves the range and operating efficiency of electric tractors, reduces RMSE prediction errors by more than 50%, and is suitable for multiple working condition switching scenarios.
Smart Images
Figure CN122153421B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electric agricultural machinery power prediction technology, specifically to an online power prediction method for electric tractors based on operating condition separation. Background Technology
[0002] With the continuous improvement of agricultural mechanization, informatization, and intelligentization, the application of new energy electric tractors in the agricultural field is becoming increasingly common. However, compared with traditional fuel tractors, new energy electric tractors suffer from insufficient power and poor range during operation. The fundamental reason lies in the limitations of battery technology. In complex field environments, the tractor's workload is affected by various factors such as soil conditions and operation type, resulting in significant nonlinear and random fluctuations in its power demand. This leads to large online prediction errors, further exacerbating energy dispatch deviations and reducing the tractor's range.
[0003] Existing methods for predicting tractor power mostly rely on historical power data, employing neural networks, support vector machines, and other techniques to build prediction models. While these methods can achieve some predictive accuracy, tractors often operate under various conditions, such as plowing, rotary tilling, and driving, with switching between these conditions significantly altering the statistical characteristics and patterns of power series variations. Using a uniform model to predict power under all conditions makes it difficult to account for the power variation characteristics under different operating conditions.
[0004] Some researchers have proposed the idea of segmented prediction based on different operating conditions, which involves dividing samples into different operating conditions and calling the corresponding prediction model based on the identification results. While this approach does significantly improve prediction accuracy, it directly uses the online operating condition identification results as discrete category labels and selects a single operating condition model for prediction based on these results. This method is essentially a hard-switching approach, lacking continuous characterization of the current sample's operating condition affiliation and the characteristics of operating condition transitions. When a sample is near the operating condition boundary or during an operating condition transition, if the online identification results are lagging, fluctuating, or misjudged, it can easily lead to inaccurate selection of the prediction model, thereby increasing the online prediction error. Furthermore, most existing technologies focus on the operating condition category discrimination itself, without further characterizing the relative proximity between the current sample and the centers of each operating condition, failing to effectively distinguish whether the sample is in the steady-state region or the operating condition transition region. This results in the use of the same prediction strategy in both the stable and transitional operating phases, making it impossible to simultaneously ensure prediction accuracy in the steady-state region and robustness in the switching region. This limits the effectiveness of existing methods in online tractor power prediction scenarios, reduces the stability and reliability of prediction results, and leads to a decrease in the energy utilization efficiency of tractors. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an online power prediction method for electric tractors based on working condition separation. By continuously quantifying the working state through working condition separation and adaptively switching between single-model and fusion model prediction strategies, this method solves the technical problems of large prediction errors and poor stability during working condition switching.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for online prediction of the power of an electric tractor based on the separation of operating conditions includes the following steps:
[0008] Step S1: In the offline phase, the collected tractor operating power data is divided into samples using a sliding window method, and power feature parameters that can characterize the operating status are extracted; the specific steps are as follows:
[0009] Step S1.1, the offline power sequence is processed using a length of... The sample is partitioned using a sliding window, and the window update step size is set to 1. Each window contains data representing a sample.
[0010] Step S1.2: Extract feature parameters that characterize the power characteristics from the power sequence within the sliding window. The extracted feature parameters include indicators such as average power, power standard deviation, maximum power, power change rate, and power peak factor. Construct sample feature vectors based on these indicators.
[0011] Step S1.3: Perform Principal Component Analysis (PCA) on the sample feature vectors to reduce dimensionality and obtain a low-dimensional sample feature set.
[0012] Step S2 involves first reducing the dimensionality of the sample features extracted in Step S1, then performing work condition clustering to obtain the work condition centers and datasets for typical operations, and finally training the corresponding power prediction model based on each work condition dataset. The specific steps are as follows:
[0013] Step S2.1: Perform cluster analysis using the dimensionality-reduced sample features. Based on the cumulative variance contribution rate criterion, select the top principal components corresponding to the cumulative variance contribution rate reaching 90% as the dimensionality-reduced sample features for working condition clustering to obtain the typical working conditions of the tractor and its working condition center.
[0014] Step S2.2: Based on the clustering results, obtain the dataset of typical working conditions, and construct corresponding power prediction models for different working conditions.
[0015] Step S2.3: When constructing the power prediction model, the power data at the current time and several past times are used as input, and the power data at future times are used as output to train each power prediction model, thereby obtaining multiple power prediction models.
[0016] Step S3: For the power data within the current sliding window, first extract the same feature parameters as in step S1, then perform dimensionality reduction using the same method as in step S2, and finally calculate the distance relationship between the current main operating condition and the center of each typical operating condition to obtain the separation degree between the current main operating condition and the real-time operating condition. The specific steps are as follows:
[0017] Step S3.1, during tractor operation Based on the power at the current and past times, a low-dimensional feature vector is obtained through feature parameter extraction and PCA dimensionality reduction. ,calculate With each working condition center The Euclidean distance between them is calculated using the following formula:
[0018] (1)
[0019] Step S3.2: The working condition with the smallest distance to the tractor's current characteristic is recorded as the main working condition, and its distance is denoted as... The working condition with the second smallest distance is denoted as the secondary primary working condition, and its distance is denoted as... The relevant formulas are as follows:
[0020] (2)
[0021] (3)
[0022] Step S3.3: Construct the working condition separation degree based on the distance relationship between the primary working condition and the secondary primary working condition.
[0023] (4)
[0024] in, It is a very small positive number used to suppress the abnormal amplification of the working condition separation caused by the denominator being too small, and to scale the index, thereby improving stability and noise resistance.
[0025] Step S3.4: Output the separation degree between the main working condition and the working condition based on the distance from the sample to the center point of each working condition.
[0026] Step S4, online power prediction stage, according to The relationship with the operating condition separation threshold determines whether the tractor's current operating state is in the steady-state region or the switching region. Based on the determination result, different prediction strategies are used to obtain the final power prediction output. The specific steps are as follows:
[0027] Step S4.1, using the working condition separation threshold With the maximum fusion ratio coefficient As parameters to be optimized, with the goal of minimizing the prediction error of offline power data, a composite function is constructed as shown in equation (5):
[0028] (5)
[0029] in, Total sample size; This is the actual power value; For a given and Power prediction under the given conditions; The working condition separation threshold determines whether the tractor is currently in the steady-state zone or the switching zone. The maximum fusion ratio coefficient represents the degree to which non-primary load case models participate in prediction; is the regularization coefficient, used to balance the prediction error with the degree to which the parameter to be determined deviates from the prior value; This is the prior value of the working condition separation threshold; This is the prior value of the maximum fusion ratio coefficient; for The prior variance is used to Measuring allowable deviation from the prior center The range of values. The smaller the value, the better. The stronger the penalty for deviating from the prior value; for The prior variance is used to measure Allowing deviation from the prior center The range of values. The smaller the value, the better. The stronger the penalty for deviating from the prior value.
[0030] Equation (5) establishes a unified optimization criterion between minimizing prediction error and parameter stability constraints, through the optimization of... and By performing joint optimization, we can obtain the optimal parameter combination that balances prediction accuracy and parameter rationality, and avoid problems such as sensitive model switching, unbalanced fusion output, or increased fluctuations in prediction results caused by excessive deviation of parameter values from the prior range.
[0031] Step S4.2: Based on the distribution range of operating condition separation in the offline power data and the perturbation constraint of the fused output on the prediction results, determine the operating condition separation thresholds respectively. With the maximum fusion ratio coefficient Prior values and Further, based on the sample separation distribution characteristics and the allowable participation level of the non-master case model, the allowable fluctuation range of both is determined, and the parameter search interval is given. The optimization process is as shown in equation (6):
[0032] (6)
[0033] in, and The optimal value is yet to be determined.
[0034] Step S4.3: Set the operating condition separation threshold according to step S4.2. To prevent the model from oscillating at the boundary between the operating condition switching region and the steady-state region, let... Let be the threshold for the tractor to enter the switching region from the steady-state region. , which is the threshold for the tractor to enter the steady-state region from the working condition switching zone.
[0035] Step S4.4: In the initial stage of power prediction, since the working condition identification window length is greater than the power prediction window length, the current state cannot be updated when the tractor's running time has not reached the working condition identification window length in the early stage of tractor operation. Therefore, the power prediction model trained by all training samples is used to predict the power. When the working condition identification window length requirement is met for the first time, the current working condition state is updated and the working condition at that moment is used as the initial working condition.
[0036] Step S4.5: Based on the separation degree between the operating condition at the previous moment and the current moment, determine whether the current operating condition is in the steady-state region or the operating condition transition region; when When the tractor is in a steady-state operating condition, only the prediction model corresponding to the main operating condition is used for power prediction; when When the current state is determined to be in the working condition switching zone, the prediction results of the corresponding models for the primary working condition and the secondary primary working condition are weighted and fused; when If the current operation status is consistent with the previous status, then the current operation status remains the same.
[0037] Step S4.6: Adaptively adjust the fusion ratio coefficient based on the working condition separation index. When the tractor is in the steady-state working condition zone, The prediction model uses a single model initially; however, when in a transitional operating condition zone, the prediction model uses multi-model fusion. according to Adaptive adjustment, calculated as shown in equation (7); due to the use of single-model prediction versus fusion model prediction and Therefore, the unified power prediction model is as shown in equation (8):
[0038] (7)
[0039] (8)
[0040] Equation (7) is used to adaptively adjust the model fusion ratio based on the separation degree of the working conditions. Equation (8), and These are the prediction outputs of the primary operating condition model and the secondary primary operating condition model, respectively. When the tractor's operating state is within the steady-state region, The value is 0; when the tractor is in the working condition switching zone, Adaptive adjustment based on operating condition separation index; power prediction output based on... The participation levels of the primary operating condition model and the secondary primary operating condition model are adaptively adjusted to reduce the prediction error in the switching zone.
[0041] Furthermore, the typical operating conditions of the electric tractor collected in step S1 include plowing, rotary tilling, and road driving conditions.
[0042] Furthermore, the power prediction model constructed in step S2 can be any one of radial basis function neural network, LSTM neural network, GRU neural network or support vector machine.
[0043] Furthermore, in step S2.1, the K-Means clustering algorithm is used to perform working condition clustering on the dimensionality-reduced sample features.
[0044] Furthermore, the final power prediction value output in step S4 is input in real time to the vehicle energy management system of the electric tractor for power battery charging and discharging power scheduling, drive motor torque feedforward control, power matching adjustment of the work implements, and real-time estimation of remaining working time and driving range.
[0045] This technology proposes an online power prediction method for electric tractors with working condition separation characteristics, which has the following advantages and beneficial effects:
[0046] This invention introduces a working condition separation index to continuously quantify the degree of belonging of real-time samples to each working condition, replacing the traditional discrete working condition identification, accurately distinguishing between steady-state areas and switching areas, and eliminating the mismatch problem caused by hard model switching.
[0047] An adaptive fusion prediction strategy is adopted. In the steady-state region, a single model ensures prediction accuracy, while in the switching region, multiple models are weighted and fused to suppress error fluctuations. The RMSE prediction error is reduced by more than 50%, balancing prediction accuracy and robustness.
[0048] Combining PCA dimensionality reduction and K-Means clustering reduces computational complexity and adapts to the real-time computing requirements of electric tractor onboard controllers; joint parameter optimization combined with prior constraints avoids prediction oscillations caused by parameter imbalance and improves method stability.
[0049] This invention can accurately adapt to multiple working conditions such as plowing, rotary tilling, and driving, providing reliable real-time power data for energy management and power scheduling of electric tractors, effectively improving the equipment's endurance and operating efficiency. Attached Figure Description
[0050] Figure 1 This is a fusion framework diagram of the power prediction model representing the working condition separation degree of the present invention;
[0051] Figure 2 This is a timing diagram of the offline power data of the electric tractor in an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram of the variance contribution rate of each principal component after PCA dimensionality reduction in an embodiment of the present invention;
[0053] Figure 4 This is a comparison chart of the prediction errors of the single-condition model and the fusion model of this invention;
[0054] Figure 5 for Figure 4 Comparison chart of predicted and actual power values for section a in the middle;
[0055] Figure 6 for Figure 4 Comparison of predicted and actual power values for section d in the middle. Detailed Implementation
[0056] The present invention will be further described below with reference to the embodiments. It should be noted that the following content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should all be considered to fall within the protection scope of the present invention.
[0057] Please see Figures 1-6 As shown, this invention provides an online power prediction method for electric tractors based on operating condition separation, comprising the following steps:
[0058] Step S1: Divide the offline power data into segments using a sliding window method and extract feature parameters; the specific steps are as follows:
[0059] Step S1.1: Collect historical power data of the tractor under three typical operating conditions: plowing, rotary tilling, and driving, totaling 10197 seconds. The offline power time series is divided into samples using a sliding window of 50 seconds, with a window update step size of 10 seconds. The data within each window represents one sample, indicating the tractor's operating status during that time period. The collected offline power data is as follows: Figure 2 ;
[0060] Step S1.2 involves extracting statistical features that characterize power fluctuations from the power sequence within the sliding window, constructing a sample feature vector, and extracting features including average power, power standard deviation, maximum power, power change rate, and power peak factor. The physical meaning of each sample will be characterized by these feature parameters. However, directly using these feature indicators for cluster analysis not only significantly increases the computational load but also introduces redundant information between features, thus requiring dimensionality reduction.
[0061] Step S1.3 involves performing Principal Component Analysis (PCA) to reduce the dimensionality of the extracted feature parameters. Since the extracted feature parameters contain some redundancy, and directly performing working condition clustering would significantly increase the computational load on the tractor's internal controller, PCA is used to reduce the dimensionality of the sample feature vectors and improve the efficiency of subsequent clustering calculations. This yields a low-dimensional feature sample set. Based on the cumulative variance contribution rate criterion, and according to well-known principles in this field, a cumulative variance contribution rate of 90% for the principal components after dimensionality reduction can effectively interpret the original data information. The variance contribution rates of each principal component after dimensionality reduction are shown below. Figure 3 As shown in the figure. The dimensionality reduction results indicate that the cumulative variance contribution rate reaches 90%, which includes the first 5 principal components. Therefore, the subsequent cluster analysis will use these five principal components.
[0062] Step S2: Perform operating condition clustering on the samples, and use the obtained operating condition dataset to construct power prediction models for different operating conditions; the specific steps are as follows:
[0063] Step S2.1: Use the K-Means algorithm to perform condition clustering on the dimensionality-reduced feature samples. To match the collected tractor operating condition categories, set the clustering parameter in the K-Means clustering algorithm to 3. According to step S1.3, use the first five principal components as the dimensionality-reduced samples for condition clustering to obtain sample data and condition centers under different tractor operating conditions.
[0064] Step S2.2: Based on the clustering results, the historical power data is divided into three typical working condition datasets: driving, plowing, and rotary tilling. For each working condition, a corresponding power prediction model is constructed using a Radial Basis Function (RBF) neural network. In this embodiment, the power prediction model uses an RBF neural network; however, the method of this invention is not limited to RBF neural networks and is also applicable to other time series prediction models.
[0065] Step S2.3: When training the RBF neural network, the power data of the current time and the past 10 seconds are used as input, and the power data of the next 3 seconds are used as output. The corresponding RBF power prediction model is trained according to the characteristics of different working conditions.
[0066] Step S3: Calculate the working condition separation degree of the current sample based on the distance between the real-time sample and the center of each working condition; the specific steps are as follows:
[0067] Step S3: Extract the same feature parameters as in Step S1 from the power data within the current sliding window, then perform feature dimensionality reduction using the same dimensionality reduction method as in Step S2. Finally, calculate the distance relationship between the power data and the centers of each typical operating condition to obtain the separation degree between the current main operating condition and the real-time operating condition. The specific steps are as follows:
[0068] Step S3.1, during tractor operation Based on the power at the current and past times, a low-dimensional feature vector is obtained through feature parameter extraction and PCA dimensionality reduction. ,calculate With each working condition center The Euclidean distance between them is calculated using the following formula:
[0069] (1)
[0070] Step S3.2: The working condition with the smallest distance to the tractor's current characteristic is recorded as the main working condition, and its distance is denoted as... The working condition with the second smallest distance is denoted as the secondary primary working condition, and its distance is denoted as... The relevant formulas are as follows:
[0071] (2)
[0072] (3)
[0073] Step S3.3: Construct the working condition separation degree based on the distance relationship between the sample and the main working condition and the secondary main working condition.
[0074] (4)
[0075] in, It is a very small positive number used to suppress the abnormal amplification of the working condition separation caused by the denominator being too small, and at the same time, the index is scaled to improve stability and noise resistance.
[0076] Step S3.4: Based on the distance from the sample to the center point of each working condition, output the separation degree between the main working condition and the working condition. When, it indicates that the current tractor operation is in the steady-state range; when The time indicates that the tractor is currently in a switching zone between different operating conditions; when At this time, the tractor's operating state remains consistent with the previous state. If the previous state was in the steady-state zone, the current state is still in the steady-state zone; if the previous state was in the switching zone, the current state is still in the switching zone.
[0077] Step S4, the online power prediction stage, first compare the real-time operating condition separation. The relationship with the working condition separation threshold determines whether to use a single model or a fusion model for prediction, and then... Determine the real-time fusion ratio coefficient Finally, the predicted power is output; the specific steps are as follows:
[0078] Step S4.1, using the working condition separation threshold With the maximum fusion ratio coefficient As parameters to be optimized, with the goal of minimizing the prediction error of offline power data, a composite function is constructed as shown in equation (5):
[0079] (5)
[0080] in, Total sample size; This is the actual power value; For a given and Power prediction under the given conditions; The working condition separation threshold determines whether the tractor is currently in the steady-state zone or the switching zone. The maximum fusion ratio coefficient represents the degree to which non-primary load case models participate in prediction; is the regularization coefficient, used to balance the prediction error with the degree to which the parameter to be determined deviates from the prior value; This is the prior value of the working condition separation threshold; This is the prior value of the maximum fusion ratio coefficient; for The prior variance is used to measure Allowing deviation from the prior center The range of values. The smaller the value, the better. The stronger the penalty for deviating from the prior value; for The prior variance is used to measure Allowing deviation from the prior center The range of values. The smaller the value, the better. The stronger the penalty for deviating from the prior value.
[0081] Equation (5) establishes a unified optimization criterion between minimizing prediction error and parameter stability constraints, through the optimization of... and By performing joint optimization, we can obtain the optimal parameter combination that balances prediction accuracy and parameter rationality, and avoid problems such as sensitive model switching, unbalanced fusion output, or increased fluctuations in prediction results caused by excessive deviation of parameter values from the prior range.
[0082] Step S4.2: Based on the distribution range of operating condition separation in the offline power data and the perturbation constraint of the fused output on the prediction results, determine the operating condition separation thresholds respectively. With the maximum fusion ratio coefficient Prior values and Further, based on the sample separation distribution characteristics and the allowable participation level of the non-master case model, the allowable fluctuation range of both is determined, and the parameter search interval is given. The optimization process is as shown in equation (6):
[0083] (6)
[0084] in, and The optimal value is yet to be determined.
[0085] Step S4.3: Set the operating condition separation threshold according to step S4.2. To prevent the model from oscillating at the boundary between the operating condition switching region and the steady-state region, let... , is the threshold for the tractor to enter the switching zone from the steady-state operating region, let , which is the threshold for the tractor to enter the steady-state region from the working condition switching zone.
[0086] Step S4.4: In the initial stage of power prediction, since the working condition identification window length is greater than the power prediction window length, working condition identification cannot be performed when the tractor running time has not reached the working condition identification window length in the early stage of tractor operation. Therefore, the power prediction model trained on all offline samples is used to predict the power. When the working condition identification window length requirement is met for the first time, the current working condition status is updated, the corresponding power prediction model is called to predict the power, and the working condition at that moment is used as the initial working condition.
[0087] Step S4.5: Based on the separation degree between the operating condition at the previous moment and the current moment, determine whether the current operating condition is in the steady-state region or the switching region; when When the tractor is in a steady-state operating condition, only the prediction model corresponding to the main operating condition is used for power prediction; when When the current state is determined to be in the working condition switching zone, the prediction results of the corresponding models for the primary working condition and the secondary primary working condition are weighted and fused; when If the current operation status is consistent with the previous status, then the current operation status remains the same.
[0088] Step S4.6: Adaptively adjust the fusion ratio coefficient based on the working condition separation index. When the tractor is in the steady-state working condition zone, The prediction model uses a single model initially; however, when in a transitional operating condition zone, the prediction model uses multi-model fusion. according to Adaptive adjustment, calculated as shown in equation (7); due to the use of single-model prediction versus fusion model prediction and Therefore, the unified power prediction model is as shown in equation (8):
[0089] (7)
[0090] (8)
[0091] Equation (7) adaptively adjusts the model fusion scaling factor based on the working condition separation degree. In Equation (8), and These are the prediction outputs of the primary operating condition model and the secondary primary operating condition model, respectively. The output values are based on... The participation levels of the primary operating condition model and the secondary primary operating condition model are adjusted to reduce prediction errors caused by operating condition identification errors during the switching zone. To verify the effectiveness of the proposed scheme, operating condition-based sub-operating condition power prediction is used as a comparison scheme. The prediction errors of the two schemes are compared as follows: Figure 4 , Figure 4 The yellow shaded area represents the operating condition switching zone. The numbers a to l in the figure are the sequence numbers of each switching zone, used to distinguish switching events within different time periods, facilitating subsequent comparative analysis of prediction errors and model performance in each interval. Regarding... Figure 4 For two typical sections, a and d, a comparison between the predicted and actual values of the two schemes is given. Figure 5 , Figure 6 .
[0092] according to Figure 4 The results show that, compared with the single-model prediction method based on working condition identification, the overall prediction error curve of the fusion model method is flatter, and the error peak is significantly suppressed. Quantitative results show that the RMSE of the fusion model prediction results decreased from 9.8717 in the single-model prediction to 5.2470, and the prediction error decreased by 46.8%, indicating that the proposed method can significantly reduce the overall prediction error. Specifically, this is reflected in... Figure 4In the a~l section, due to the potential similarities in characteristics among typical work condition segments and the dynamic nature of work condition switching, the window contains features from both the previous and subsequent work conditions. This leads to significant transitional and ambiguous features in feature extraction, resulting in identification errors and significantly increasing prediction errors when incorrect models are called. The proposed method, however, adaptively adjusts the model fusion ratio based on the work condition separation degree, mitigating the model mismatch problem caused by hard switching in work condition identification. It exhibits more stable suppression of error fluctuations and improves prediction accuracy during the transition phase. Figure 4 In the diagram, point 'a' corresponds to a typical operational transition process, representing the difference in dynamic tracking performance between the two methods within a general switching zone. Point 'd' corresponds to the maximum error peak value across the entire time domain, reflecting the predictive capability of the method under extreme mismatch scenarios. Therefore, for... Figure 4 Further analysis will be conducted on the two typical cases at points a and d. It should be noted that... Figure 5 and Figure 6 It's not about Figure 4 Instead of simply magnifying the error curves at points a and d, this paper presents a comparison between the predicted and actual values for the two representative segments a and d.
[0093] according to Figure 5 The results show that in the steady-state region of the initial operating condition, both methods demonstrate good prediction accuracy and tracking ability of the actual values. However, during the transition phase where power rapidly decreases and then surges again, the single-model predictions begin to show significant local fluctuations and tracking deviations, indicating a potential model switching mismatch in this segment, leading to increased prediction errors in the single models. The fusion model, on the other hand, effectively follows the changing trend of the actual power. Figure 6 The results demonstrated not only exhibit prediction errors present in the switching conditions region, but also highlight the significant power surges and short-duration spikes that follow. This leads to pronounced peak mismatch and dynamic lag in single-model predictions. Under continuous power surges, single-model prediction methods are more prone to error amplification due to model mismatch. While the fusion model prediction still exhibits some bias in this region, its tracking of the actual power change trend is significantly superior to the single model, especially near transitions and peaks, where it better approximates the true curve. Analysis of the power prediction results under these two typical conditions shows that the proposed scheme can effectively mitigate the error amplification caused by model mismatch in complex switching regions.
[0094] Overall, the online power prediction method based on the operating condition separation degree can guarantee the prediction accuracy in the steady-state region of the operating condition and effectively suppress the prediction error in the operating condition switching region, indicating that the proposed method has good applicability and effectiveness under complex operating condition switching conditions.
[0095] The above is an exemplary description of the invention. Obviously, the specific implementation of the invention is not limited to the above-described manner. Any non-substantial improvement made using the inventive concept and technical solution of the invention, or the direct application of the inventive concept and technical solution to other situations without modification, is within the protection scope of the invention.
Claims
1. A method for online prediction of the power of an electric tractor based on working condition separation, characterized in that, Includes the following steps: Step S1: Offline data processing: Collect typical operating power data of electric tractors, use a sliding window to divide the power sequence into samples, extract power feature parameters and perform PCA dimensionality reduction to obtain a low-dimensional sample feature set. Step S2, Construction of working condition model: Perform working condition clustering on the dimensionality-reduced sample features to obtain the typical working condition center and each working condition dataset, and train the corresponding power prediction model based on each working condition dataset. Step S3: Real-time operating condition separation calculation: Extract the feature parameters of the power data in the current sliding window and reduce the dimensionality, calculate the Euclidean distance between it and the center of each operating condition, determine the main operating condition and the secondary main operating condition, and construct and calculate the real-time operating condition separation based on the distance relationship. Step S4, Adaptive online power prediction: Preset the operating condition separation threshold, and determine the operating status as the steady state or the operating condition switching zone based on the relationship between the real-time operating condition separation and the threshold; the steady state zone uses the main operating condition single model prediction, and the switching zone uses the weighted fusion prediction of the main and secondary operating condition models, adaptively adjusts the fusion ratio coefficient, and outputs the final power prediction value. Step S3 specifically includes the following steps: Step S3.1, during tractor operation Based on the power at the current and past times, a low-dimensional feature vector is obtained through feature parameter extraction and PCA dimensionality reduction. ,calculate With each working condition center The Euclidean distance between them is calculated using the following formula: (1)； Step S3.2: The working condition with the smallest distance to the tractor's current characteristic is recorded as the main working condition, and its distance is denoted as... The working condition with the second smallest distance is denoted as the secondary primary working condition, and its distance is denoted as... The relevant formulas are as follows: (2) (3)； Step S3.3: Construct the working condition separation degree based on the distance relationship between the primary working condition and the secondary primary working condition. (4) in, It is a positive number, used to suppress the abnormal amplification of the working condition separation caused by an excessively small denominator, and to perform scale compression on the index; Step S3.4: Output the separation degree between the main working condition and the working condition based on the distance from the sample to the center point of each working condition.
2. The method for online prediction of electric tractor power based on working condition separation as described in claim 1, characterized in that, Step S1 specifically includes the following steps: Step S1.1, the offline power sequence is processed using a length of... The sample is partitioned using a sliding window, and the window update step size is set to 1. Each window contains data representing a sample; Step S1.2: Extract feature parameters that characterize the power characteristics from the power sequence within the sliding window, and construct a sample feature vector based on these feature parameters; Step S1.3: Perform principal component analysis to reduce the dimensionality of the sample feature vectors, and obtain a low-dimensional sample feature set.
3. The method for online prediction of electric tractor power based on working condition separation as described in claim 2, characterized in that, The characteristic parameters include average power, power standard deviation, maximum power, power change rate, and power peak factor.
4. The method for online prediction of electric tractor power based on working condition separation as described in claim 1, characterized in that, Step S2 specifically includes the following steps: Step S2.1: Perform cluster analysis using the dimensionality-reduced sample features, select the principal components corresponding to the cumulative variance contribution rate reaching 90% as the dimensionality-reduced sample features for working condition clustering, and obtain the typical working conditions of the tractor and its working condition center. Step S2.2: Based on the clustering results, obtain the dataset of typical working conditions, and construct corresponding power prediction models for different working conditions. Step S2.3: When constructing the power prediction model, power data at the current time and several past times are used as input, and power data at future times are used as output to construct various power prediction models, resulting in multiple power prediction models.
5. The method for online prediction of power of an electric tractor based on working condition separation as described in claim 1, characterized in that, Step S4 specifically includes the following steps: Step S4.1, using the working condition separation threshold With the maximum fusion ratio coefficient As parameters to be optimized, with the goal of minimizing the prediction error of offline power data, a composite function is constructed as shown in equation (5): (5) in, Total sample size; This is the actual power value; For a given and Power prediction under the given conditions; The working condition separation threshold; This is the maximum fusion ratio coefficient; The regularization coefficient is used. This is the prior value of the working condition separation threshold; This is the prior value of the maximum fusion ratio coefficient; for The prior variance; for The prior variance; Step S4.2: Based on the distribution range of the operating condition separation degree in the offline power data and the perturbation constraint of the fused output on the prediction results, determine the operating condition separation degree threshold respectively. With the maximum fusion ratio coefficient Prior values and Further, based on the sample separation distribution characteristics and the allowable participation level of the non-master case model, the allowable fluctuation range of both is determined, and the parameter search interval is given; the optimization process is as shown in equation (6): (6) in, and The optimal value is yet to be determined; Step S4.3: Set the operating condition separation threshold according to step S4.
2. To prevent the model from oscillating at the boundary between the operating condition switching region and the steady-state region, let... Let be the threshold for the tractor to enter the switching region from the steady-state region. , is the threshold for the tractor to enter the steady-state region from the working condition switching zone; Step S4.4: In the initial stage of power prediction, the power prediction model trained with all training samples is used to predict the power. When the working condition identification window length requirement is met for the first time, the current working condition state is updated, and the working condition at the moment when the working condition identification window length requirement is met for the first time is used as the initial working condition. Step S4.5: Based on the separation degree between the operating condition at the previous moment and the current moment, determine whether the current operating condition is in the steady-state region or the operating condition transition region; when When the tractor is in a steady-state operating condition, then its operating condition is within the steady-state range; when When the current state is determined to be in the working condition switching zone, the prediction results of the corresponding models for the primary working condition and the secondary primary working condition are weighted and fused; when If the current operation status remains consistent with the previous status, then the current operation status will remain consistent with the previous status. Step S4.6: Adaptively adjust the fusion ratio coefficient based on the working condition separation index. When the tractor is in the steady-state working condition zone, The prediction model uses a single model initially; however, when in a transitional operating condition zone, the prediction model uses multi-model fusion. according to Adaptive adjustment, calculated as shown in equation (7); since both single-model prediction and fusion model prediction are related to... The unified power prediction model is as shown in equation (8): (7) (8)。 6. The method for online prediction of electric tractor power based on working condition separation as described in claim 1, characterized in that, The typical operating conditions of the electric tractor collected in step S1 include plowing, rotary tilling, and road driving.
7. The method for online prediction of electric tractor power based on working condition separation as described in claim 1, characterized in that, The power prediction model constructed in step S2 can be any one of radial basis function neural network, LSTM neural network, GRU neural network or support vector machine.
8. The method for online prediction of electric tractor power based on working condition separation as described in claim 4, characterized in that, In step S2.1, the K-Means clustering algorithm is used to perform working condition clustering on the dimensionality-reduced sample features.
9. The method for online prediction of electric tractor power based on working condition separation as described in claim 1, characterized in that, The final power prediction value output in step S4 is input to the electric tractor's vehicle energy management system in real time for power battery charging and discharging power scheduling, drive motor torque feedforward control, power matching adjustment of work implements, and real-time estimation of remaining working time and driving range.