Range extender temperature prediction method and device, computer equipment and storage medium

Abstract

The invention relates to a range extender temperature prediction method and device, computer equipment and a storage medium, and the method comprises the steps: collecting time sequence data of a plurality of data sources of a range extender, and determining the contribution degree weight of each data source for the temperature prediction of the range extender; processing the time sequence data of each data source based on a preset processing method to obtain the processed time sequence data of each data source, and taking the processed time sequence data of each data source as the time sequence data of the dimension expansion data source; determining the contribution degree weight of each dimension expansion data source; constructing input features of the temperature prediction model; determining a position coding weight of a feature value of each dimension feature in the input features, and performing adaptive fusion processing on the input features according to the position coding weight of the feature value of each dimension feature to obtain a fusion feature; and inputting the fusion features into a temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model. According to the method, the accuracy and reliability of temperature prediction of the range extender can be improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a method, apparatus, computer device, and storage medium for predicting the temperature of a range extender. Background Technology

[0002] As the range requirements of electric vehicles increase, range extenders, as a key auxiliary power system, use an internal combustion engine to generate electricity to charge the battery and extend the driving range. However, range extenders generate a significant amount of heat during operation. If the range extender temperature becomes too high, it may cause abnormalities leading to equipment damage, accelerated battery aging, or even safety accidents. Therefore, real-time temperature prediction of the range extender is crucial for ensuring the safety and stability of electric vehicles.

[0003] In existing technologies, predicting the temperature of range extenders using traditional physical modeling methods typically requires detailed modeling of factors such as the internal structure and thermal conductivity of the range extender, and relies heavily on engineering experience and assumptions. Due to the complexity of the range extender's internal structure and the variability of its operating conditions, traditional physical modeling methods struggle to handle these dynamic changes, resulting in limited prediction accuracy.

[0004] With the rapid development of deep learning technology, temperature prediction methods for range extenders based on deep learning have gradually attracted attention. This method can perform deep learning modeling based on historical data of the range extender, overcoming the limitations of physical models to some extent. However, high-quality sample data suitable for model training is scarce, especially sample data under extreme operating conditions. The problem of few samples severely limits the generalization ability and prediction accuracy of deep learning models, making it difficult to meet the high-precision and high-reliability temperature prediction requirements in practical applications. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, computer equipment, and storage medium for predicting the temperature of a range extender, which can improve the accuracy and reliability of range extender temperature prediction.

[0006] A method for predicting the temperature of a range extender includes: collecting time-series data from multiple data sources of the range extender and determining the contribution weight of each data source to the temperature prediction of the range extender; processing the time-series data from each data source based on a preset processing method to obtain processed time-series data from each data source, and using the processed time-series data from each data source as time-series data from extended-dimensional data sources; determining the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source; constructing input features of a temperature prediction model based on the time-series data and contribution weights of each data source, the time-series data from each extended-dimensional data source, and the contribution weights of each extended-dimensional data source, wherein the input features include multiple dimensional features, and each dimensional feature corresponds to each data source or each extended-dimensional data source; determining the positional encoding weight of the feature values ​​of each dimensional feature in the input features; performing adaptive fusion processing on the input features according to the positional encoding weight of the feature values ​​of each dimensional feature to obtain fused features; and inputting the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0007] In one embodiment, determining the contribution weight of each data source to the range extender temperature prediction includes: collecting sample time-series data from each data source and sample temperature data of the range extender corresponding to the sample time-series data from each data source; obtaining the contribution weight of each data source based on the sample time-series data from each data source, the corresponding sample temperature data of the range extender, the temperature prediction model, and using Shapley additive interpretation.

[0008] In one embodiment, the preset processing method includes a set mathematical transformation method and / or a feature interaction processing method. The preset processing method is used to process time-series data from each data source, including: processing time-series data from a first target data source (which may be one or more data sources) based on the mathematical transformation method; and / or processing time-series data from a second and third target data source (which may be any one of multiple data sources) based on the feature interaction processing method; and determining the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source, including: if a mathematical transformation method is used to process the time-series data from the first target data source, a contribution weight processing method is determined based on the mathematical transformation method, and the contribution weight of the first target data source is processed according to the contribution weight processing method to obtain the contribution weight of the extended-dimensional data source corresponding to the first target data source; if a feature interaction processing method is used to process the time-series data from the second and third target data sources, the contribution weight of the processed extended-dimensional data source is determined based on the contribution weight of the second and third target data sources.

[0009] In one embodiment, the input features of the temperature prediction model are constructed based on the time-series data of each data source and the contribution weights of each data source, as well as the time-series data of each extended-dimensional data source and the contribution weights of each extended-dimensional data source. This includes: multiplying the time-series data of each data source with the contribution weights of each data source to obtain the feature values ​​of the dimensional features corresponding to each data source; multiplying the time-series data of each extended-dimensional data source with the contribution weights of each extended-dimensional data source to obtain the feature values ​​of the dimensional features corresponding to each extended-dimensional data source; and constructing the input features of the temperature prediction model based on the feature values ​​of the dimensional features corresponding to each data source and the dimensional features corresponding to each extended-dimensional data source.

[0010] In one embodiment, before determining the positional encoding weights of the feature values ​​of each dimension of the input features, a range extender temperature prediction method further includes: dividing the input features into time windows to obtain sub-input features for each time window, each sub-input feature containing feature values ​​corresponding to multiple consecutive time points in each time window; determining the positional encoding weights of the feature values ​​of each dimension of the input features, including: for any sub-input feature, performing positional encoding calculations on the feature values ​​of each dimension based on the parity of the index of each dimension feature, the feature value of each dimension feature, and the positional index of the feature value of each dimension feature, to obtain the positional encoding weights of the feature values ​​of each dimension feature.

[0011] In one embodiment, the input features are adaptively fused based on the positional encoding weights of the feature values ​​of each dimension to obtain fused features. This includes: calculating the similarity between any two dimensional features based on their positional encoding weights; constructing a similarity matrix of the input features based on the similarity between any two dimensional features; using the similarity matrix as the feature position adjacency matrix of a graph convolutional neural network; and performing graph convolution on the feature position adjacency matrix and the input features using a graph convolutional neural network to obtain fused features.

[0012] In one embodiment, before inputting the fused features into the temperature prediction model, a range extender temperature prediction method further includes: for any time window of sub-fused features, determining the contribution weight of each dimension feature in the sub-fused features to the range extender temperature prediction; constructing a contribution matrix based on the contribution weight of each dimension feature in the sub-fused features to the range extender temperature prediction; calculating the similarity between any two dimension features in the sub-fused features, and constructing a feature similarity matrix based on the similarity between any two dimension features in the sub-fused features; performing collaborative filtering based on the contribution matrix and the feature similarity matrix to select multiple target dimension features whose dimensional feature correlation is greater than a set threshold; and constructing fused features based on the feature values ​​of each target dimension feature.

[0013] In one embodiment, after constructing fusion features based on the feature values ​​of each target dimension feature, a range extender temperature prediction method further includes: determining the position encoding weight of each time window based on the count value of each dimension feature in the fusion features for each time window; performing global time series information fusion processing on the fusion features of each time window based on the position encoding weight of each time window to obtain target fusion features; and inputting the fusion features into a temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model, including: inputting the target fusion features into a temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0014] A range extender temperature prediction device includes: a data acquisition module for acquiring time-series data from multiple data sources of the range extender and determining the contribution weight of each data source to the range extender temperature prediction; a processing module for processing the time-series data from each data source according to a preset processing method to obtain processed time-series data from each data source, and using the processed time-series data from each data source as time-series data from extended-dimensional data sources; a first determination module for determining the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source; a first construction module for constructing input features of a temperature prediction model based on the time-series data and contribution weights of each data source, the time-series data from each extended-dimensional data source, and the contribution weights of each extended-dimensional data source, wherein the input features include multiple dimensional features, and each dimensional feature corresponds to each data source or each extended-dimensional data source; a first fusion module for determining the positional encoding weight of the feature values ​​of each dimensional feature in the input features, and performing adaptive fusion processing on the input features according to the positional encoding weight of the feature values ​​of each dimensional feature to obtain fused features; and a prediction module for inputting the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0017] The aforementioned method, apparatus, computer equipment, and storage medium for predicting the temperature of a range extender quantifies the importance of each data source in the temperature prediction process by collecting time-series data from multiple data sources and determining the contribution weight of each data source. Then, based on a preset processing method, the time-series data of the original data sources and the contribution weights of each data source are dimensionally expanded to obtain time-series data and corresponding contribution weights of expanded-dimensional data sources. This allows for the expansion of multiple expanded-dimensional data sources from the range extender's multiple data sources, determining the time-series data and contribution weights of each expanded-dimensional data source, increasing the input sample data for the temperature prediction model, and achieving feature expansion with fewer samples. Finally, based on the time-series data and contribution weights of each data source, and the time-series data and contribution weights of each expanded-dimensional data source, the input features of the temperature prediction model containing multi-dimensional features are constructed, fully exploring the potential correlation information during the range extender's operation and providing sufficient feature representation for the model. Furthermore, by determining the positional encoding weights of feature values ​​in each dimension and adaptively fusing the input features, correlated dimensional features can be effectively aggregated while irrelevant dimensional features are suppressed, thereby optimizing the representational ability of the fused features. Finally, the fused features are input into the temperature prediction model, enabling the model to more accurately capture the potential correlations between key features and features, significantly improving the generalization ability and prediction accuracy of the deep learning model. This allows the temperature prediction model to generate high-precision temperature predictions that better reflect the actual operating state of the range extender and with controllable prediction errors. The aforementioned temperature prediction method for range extenders forms a closed loop from data processing and feature optimization to model prediction, significantly improving the accuracy and reliability of range extender temperature prediction and providing strong technical support for real-time status monitoring, anomaly warning, and efficient operation and maintenance management of range extenders. Attached Figure Description

[0018] Figure 1 This is an application environment diagram of a range extender temperature prediction method in one embodiment; Figure 2 This is a flowchart illustrating a range extender temperature prediction method in one embodiment; Figure 3 This is a schematic diagram of a process for determining the contribution weight of each data source to the temperature prediction of the range extender in one embodiment. Figure 4 This is a schematic diagram illustrating the SHAP values ​​of all time-series data from various data sources in one embodiment. Figure 5 This is a schematic diagram of a process from a data source to the input features of a temperature prediction model in one embodiment. Figure 6 This is a schematic diagram of a process for adaptively fusing input features based on the positional encoding weights of feature values ​​of each dimension, as described in one embodiment. Figure 7 This is a schematic diagram illustrating a specific process of adaptive fusion in one embodiment; Figure 8 This is a schematic diagram of a process for constructing fused features using a collaborative filtering mechanism in one embodiment; Figure 9 This is a schematic diagram illustrating a specific process for collaborative filtering of dimensional features in sub-fusion features in one embodiment. Figure 10 This is a schematic diagram of a global temporal information fusion processing method in one embodiment; Figure 11 This is a structural block diagram of a range extender temperature prediction device in one embodiment; Figure 12 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] This application provides a method for predicting the temperature of a range extender, applicable to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the range extender temperature prediction system 200 is used to execute a range extender temperature prediction method according to this application. Specifically, it collects time-series data from multiple data sources of the range extender from the terminal device 100 and determines the contribution weight of each data source to the range extender temperature prediction; it processes the time-series data of each data source based on a preset processing method to obtain processed time-series data of each data source, and uses the processed time-series data of each data source as time-series data of extended-dimensional data sources; it determines the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source; it constructs the input features of the temperature prediction model based on the time-series data of each data source and the contribution weight of each data source, the time-series data of each extended-dimensional data source and the contribution weight of each extended-dimensional data source, the input features include multiple dimensional features, each dimensional feature corresponds to each data source or each extended-dimensional data source; it determines the positional encoding weight of the feature values ​​of each dimensional feature in the input features, and performs adaptive fusion processing on the input features according to the positional encoding weight of the feature values ​​of each dimensional feature to obtain fused features; it inputs the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model. Among them, the terminal device 100 can be a mobile terminal, such as an in-vehicle terminal.

[0021] In one embodiment, such as Figure 2As shown, a method for predicting the temperature of a range extender is provided, which is then applied to... Figure 1 Taking the range extender temperature prediction system 200 as an example, the following steps are included: S201: Collect time-series data from multiple data sources for the range extender and determine the contribution weight of each data source to the temperature prediction of the range extender.

[0022] In this embodiment, the range extender temperature prediction system utilizes a multi-type sensor acquisition module to collect time-series data from multiple data sources under different operating conditions (such as start-up, idling, high load, low load, start-stop switching, etc.). Exemplarily, the collected multiple data sources may include core parameter data sources related to temperature changes during range extender operation, such as vehicle speed (related to vehicle movement), torque and speed (reflecting the range extender's own power output), current and voltage (reflecting the range extender's own electrical characteristics), and external environmental influence parameters such as ambient temperature. It is understood that in this embodiment, the collected multiple data sources may include time-series data from more or fewer data sources, and this embodiment is not limited in this respect.

[0023] Furthermore, based on the collected multi-source time-series data and the actual temperature data of the range extender in the corresponding time period, interpretability analysis methods (such as Shapley additive interpretation) are used to quantify the influence of each data source on the range extender temperature prediction results. Finally, the contribution weight of each data source is determined to intuitively reflect the importance of each data source in the range extender temperature prediction, laying the foundation for subsequent data processing and feature construction.

[0024] S202, Process the time series data of each data source based on the preset processing method to obtain the processed time series data of each data source, and use the processed time series data of each data source as the time series data of the extended dimension data source.

[0025] In this embodiment, a preset processing method is used to expand the dimensions of time-series data from each data source to uncover potential correlations and deep features, generating expanded-dimensional time-series data. For example, this preset processing method may include mathematical transformation operations (such as power transformation, logarithmic transformation, trigonometric function transformation, etc., on the original time-series data to highlight the variation patterns of the data in different numerical ranges) and feature interaction operations (such as multiplying time-series data from different data sources to construct time-series data with new feature dimensions reflecting the synergistic effects of multiple parameters). By processing the time-series data from each data source using the above preset processing method, processed time-series data from each data source are obtained and used as the expanded-dimensional time-series data for each data source, realizing the expansion of the feature dimensions of the original data and providing richer information support for subsequent model training.

[0026] S203, based on the preset processing method and the contribution weight of each data source, determine the contribution weight of each extended dimension data source.

[0027] In this embodiment, to ensure that the contribution weights of the expanded data sources are logically consistent with the weights of the original data sources and can accurately reflect the importance of the expanded features to temperature prediction, based on the preset processing method used in S202, the contribution weights of each original data source obtained in S201 are processed accordingly to generate the contribution weights of each expanded data source.

[0028] Specifically, if the preset processing method is a mathematical transformation operation, the same mathematical transformation is performed on the contribution weights of the original data source (e.g., the weights of the original data source are transformed by a power factor, and the weights of the extended data source generated by the corresponding mathematical transformation are also calculated by the same power factor); if the preset processing method is a feature interaction operation, the contribution weights of the multiple original data sources participating in the interaction are correlated (e.g., the weights of the extended features generated by the product interaction of time series data from multiple data sources can be determined by the product of the weights of the corresponding data sources), thereby determining the contribution weights of each extended data source.

[0029] By using the above method, the contribution weight of each expanded dimension data source can both inherit the importance attributes of the original data source and adapt to the uniqueness of the expanded dimension features, ensuring that the importance of features in different dimensions can be accurately quantified in the subsequent feature construction process.

[0030] S204, The input features of the temperature prediction model are constructed based on the time series data of each data source and the contribution weight of each data source, as well as the time series data of each extended dimension data source and the contribution weight of each extended dimension data source.

[0031] The input features include multiple dimensional features, each corresponding to a different data source or an extended dimensional data source.

[0032] In this embodiment, the input features of the temperature prediction model are constructed based on the time-series data and contribution weights of each data source, as well as the time-series data and contribution weights of each extended-dimensional data source. Specifically, the constructed input features include multiple dimensional features. The feature value of each dimensional feature can be calculated by weighting the time-series data of the data source with the contribution weight of the corresponding data source, or the time-series data of the extended-dimensional data source with the contribution weight of the corresponding extended-dimensional data source. This weighting method can enhance the impact of high-contribution data sources or high-contribution extended-dimensional data sources on the temperature prediction of the range extender, while suppressing the interference of noise information in low-contribution data sources or low-contribution extended-dimensional data sources on the model.

[0033] S205, determine the positional encoding weights of the feature values ​​of each dimension of the input features, and perform adaptive fusion processing on the input features according to the positional encoding weights of the feature values ​​of each dimension to obtain fused features.

[0034] In this embodiment, positional encoding technology is employed to determine the positional encoding weights of the feature values ​​of each dimension of the input features, combining the attributes of each dimension (such as dimension index, feature value magnitude, temporal position, etc.). These positional encoding weights can quantify the correlation between different dimensional features in spatial location and / or temporal order, compensating for the loss of feature associations caused by neglecting positional information in traditional feature processing. For example, for dimensional features with spatial attribute differences, positional encoding can distinguish their different roles in temperature prediction; for temporally related dimensional features, positional encoding can highlight their sequential influence over time.

[0035] Furthermore, based on the obtained positional encoding weights of the feature values ​​of each dimension, the input features are adaptively fused according to the positional encoding weights of the feature values ​​of each dimension to obtain fused features. Specifically, the fusion weights of different dimension features can be dynamically adjusted according to the feature correlation reflected by the positional encoding weights to achieve adaptive fusion of input features. For dimension features where the positional encoding weights indicate strong correlation, information aggregation is achieved through weight enhancement; for dimension features where the positional encoding weights indicate weak correlation or irrelevance, redundant information interference is suppressed through weight. The embodiments of this application can use graph convolutional neural networks for adaptive fusion to ultimately generate fused features that can effectively integrate key information and eliminate redundant interference, thereby improving the feature's ability to represent the temperature change law of the range extender.

[0036] S206, the fusion features are input into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0037] In this embodiment, the fused features obtained in S205 are input into a pre-trained temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model. This temperature prediction model can employ a deep learning model or a machine learning model, including but not limited to deep neural networks (DNNs) and graph convolutional neural networks (GCNs). The specific model selection can be based on factors such as the temporal characteristics of the range extender's temperature changes and the complexity of feature associations. For example, for temperature prediction scenarios with strong temporal dependencies, models such as Long Short Term Memory (LSTM) networks, which can capture long-term temporal dependencies, can be prioritized; for scenarios with complex feature associations and graph structure relationships, GCN models can be used to achieve deep mining of feature associations.

[0038] Furthermore, this temperature prediction model, through learning and computation of fused features, outputs the temperature prediction result of the range extender for the next time period. Further, it can perform temperature warning judgments based on the predicted temperature and the preset temperature threshold of the range extender. If the predicted temperature is higher than the preset temperature threshold, a temperature warning can be triggered, providing a basis for subsequent safety management decisions; if the predicted temperature is lower than the preset temperature threshold, it can serve as an important indicator for evaluating the operating status of the range extender, assisting in optimizing the operating parameters and maintenance strategies of the range extender.

[0039] The aforementioned method for predicting the temperature of a range extender quantifies the importance of each data source by collecting time-series data from multiple data sources and determining the contribution weight of each data source to the temperature prediction. Then, based on a pre-defined processing method, the time-series data of the original data sources and the contribution weights of each data source are dimensionally expanded to obtain time-series data and corresponding contribution weights of expanded-dimensional data sources. This allows for the expansion of multiple data sources from the range extender into multiple expanded-dimensional data sources, determining the time-series data and contribution weights of each expanded-dimensional data source, increasing the input sample data for the temperature prediction model, and achieving feature expansion with fewer samples. Finally, based on the time-series data and contribution weights of each data source, and the time-series data and contribution weights of each expanded-dimensional data source, the input features of the temperature prediction model containing multi-dimensional features are constructed, fully exploring the potential correlation information during the operation of the range extender and providing sufficient feature representation for the model. Furthermore, by determining the positional encoding weights of feature values ​​in each dimension and adaptively fusing the input features, correlated dimensional features can be effectively aggregated while irrelevant dimensional features are suppressed, thereby optimizing the representational ability of the fused features. Finally, the fused features are input into the temperature prediction model, enabling the model to more accurately capture the potential correlations between key features and features, significantly improving the generalization ability and prediction accuracy of the deep learning model. This allows the temperature prediction model to generate high-precision temperature predictions that better reflect the actual operating state of the range extender and with controllable prediction errors. The aforementioned temperature prediction method for range extenders forms a closed loop from data processing and feature optimization to model prediction, significantly improving the accuracy and reliability of range extender temperature prediction and providing strong technical support for real-time status monitoring, anomaly warning, and efficient operation and maintenance management of range extenders.

[0040] In one embodiment, such as Figure 3 As shown, determining the contribution weight of each data source to the range extender temperature prediction in step S201 above may include the following steps: S301 collects sample time-series data from each data source and the corresponding range extender sample temperature data from each data source.

[0041] In this embodiment, sample time-series data from each data source and sample temperature data of the range extender corresponding to the sample time-series data from multiple data sources are first extracted from the database of the range extender temperature monitoring system. For example, the sample time-series data and the corresponding range extender sample temperature data can be matched in the time dimension; that is, the time series lengths of the sample time-series data from each data source and the corresponding range extender sample temperature data are equal. For instance, if the sample time-series data from each data source is a time series with a sampling interval of 1 second and a total of 1000 continuous sampling points, then the range extender sample temperature data must also be a time series with the same sampling interval and the same number of sampling points, and the timestamps of each sampling point of the two types of data must correspond one-to-one.

[0042] In some embodiments, the selection of the sample temperature data for the range extender temperature needs to be aligned with the actual scenario requirements for range extender temperature prediction. For example, it can be the range extender temperature measured within the same time period corresponding to the sample time series data from each data source, or it can be the range extender temperature measured within the next time period corresponding to the time series data from each data source. The specific time period for selecting the range extender temperature data can be flexibly determined according to the actual needs of range extender temperature prediction (such as real-time monitoring, early warning, historical operating condition review, etc.) to ensure that the sample time series data collected from each data source and the sample temperature data of the range extender form a correlation that meets the prediction target, laying a data foundation for the accurate training of the subsequent initial temperature prediction model.

[0043] S302. Based on the sample time-series data of each data source and the corresponding sample temperature data and temperature prediction model of the range extender, and using Shapley additive interpretation, the contribution weight of each data source is obtained.

[0044] In this embodiment, based on sample time-series data from various data sources, corresponding range extender sample temperature data, and a pre-trained temperature prediction model, Shapley Additive Explanations (SHAP) are introduced. The S302 pre-trained temperature prediction model is used as the analysis object to quantify the contribution of each time-series data point from each data source to the range extender temperature prediction result. The pre-trained temperature prediction model can be obtained by constructing a training dataset and training the model using sample time-series data from each data source as input features and the corresponding range extender sample temperature data as output labels. It should be noted that the core function of this temperature prediction model is to establish the mapping relationship between each source data and the range extender temperature, providing a basic model carrier for subsequent quantification of the contribution of each data source, rather than being directly used for the final temperature prediction task.

[0045] Specifically, the aforementioned Shapleyness interpretation originates from game theory. Its core logic involves traversing all possible feature subsets and calculating the marginal contribution of a single feature (or data point) to the prediction result in different subset combinations. This ultimately yields a fair and interpretable contribution metric, the SHAP value, which represents the contribution of each time-series data source to the range extender temperature prediction. For example, see [link to relevant documentation]. Figure 4 ,Should Figure 4 This diagram illustrates the SHAP values ​​between time-series data from six data sources—vehicle speed, torque, engine speed, current, voltage, and ambient temperature—and the output range extender temperature. Each horizontal line represents an input feature, and the horizontal position of the point on the line corresponds to the feature's influence on the prediction of a single sample. The color of the point on the line represents the original value of that feature in that sample. From top to bottom, the horizontal lines represent torque, engine speed, current, ambient temperature, vehicle speed, and voltage. Specifically, a positive SHAP value for a given time-series data point from a given data source indicates that the data point has a positive impact on the predicted temperature; a negative SHAP value indicates that the data point has an inhibitory effect on the predicted temperature; the larger the absolute value of the SHAP value, the stronger the influence of the data point on the prediction result. By iterating through samples at all time points, the SHAP values ​​for all time-series data points under each data source can be obtained.

[0046] Furthermore, for all time-series data from each data source, the average absolute value of the SHAP values ​​is calculated to obtain the average absolute contribution value of each data source. For example, the SHAP value of any data source Fi on the j-th sample is SHAP_F. ij Where i = 1, 2, ..., 6 represents 6 input features, and j = 1, 2, ..., 1000 represents 1000 feature values. Then, calculate F for any data source. i The formula for calculating the mean absolute SHAP value is as follows: ; in, F represents the feature F on the j-th sample i The absolute value of the SHAP value, This means dividing by the total number of samples N, i.e., calculating the average.

[0047] Therefore, the average absolute contribution value of each data source calculated using the above formula can be directly used to quantify the influence of each data source on the temperature prediction of the range extender. The larger the average absolute contribution value, the stronger the average influence of the data source on the temperature prediction results across all time-series samples, and the richer the effective information related to temperature changes it contains; conversely, the smaller the average absolute contribution value, the weaker the overall influence of the data source on the temperature prediction.

[0048] Finally, the calculated average absolute contribution values ​​of each data source are normalized and mapped to a unified numerical range of [0,1], ultimately yielding the contribution weights of each data source. The core purpose of normalization is to eliminate the magnitude differences in the average absolute contribution values ​​of different data sources, ensuring the weighting results are directly comparable and additive. For example, it avoids imbalances in weight allocation caused by one data source having an excessively large average absolute contribution value (e.g., 100) while another data source has an excessively small one (e.g., 0.1). The final contribution weights, such as torque (0.3), rotational speed (0.25), current (0.2), ambient temperature (0.1), voltage (0.1), and vehicle speed (0.05), can be directly used in subsequent steps such as data expansion and feature weighting, achieving a reasonable balance between prioritizing high-contribution data sources and balancing low-contribution data sources.

[0049] In one embodiment, the preset processing method includes a set mathematical transformation method and / or a feature interaction processing method. The above-mentioned processing of time-series data from each data source based on the preset processing method to obtain processed time-series data from each data source includes: processing the time-series data of a first target data source based on the mathematical transformation method, wherein the first target data source can be one or more; and / or, processing the time-series data of a second target data source and a third target data source based on the feature interaction processing method, wherein the second target data source or the third target data source can be any one of multiple data sources; the above-mentioned determination of the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source includes: if the mathematical transformation method is used to process the time-series data of the first target data source, then a contribution weight processing method is determined based on the mathematical transformation method, and the contribution weight of the first target data source is processed according to the contribution weight processing method to obtain the contribution weight of the extended-dimensional data source corresponding to the first target data source; if the feature interaction processing method is used to process the time-series data of the second target data source and the third target data source, then the contribution weight of the processed extended-dimensional data source is determined based on the contribution weight of the second target data source and the contribution weight of the third target data source.

[0050] In this embodiment, the aforementioned preset processing method may include a set mathematical transformation method and / or feature interaction processing method. For the collected time-series data from various data sources (such as continuous time-series data like vehicle speed, torque, and RPM), mathematical transformations and / or feature interaction operations are used to process the data, resulting in processed time-series data for each data source. These processed time-series data are then used as the time-series data for each expanded-dimensional data source. Specifically, the time-series data of a first target data source is processed based on mathematical transformation methods. This first target data source can be one or more. For example, using mathematical transformation methods such as square, cube, and square root, mathematical transformations are performed on all time-series data from each data source. Based on each time-series data, three expanded-dimensional data sources are derived, and the new features maintain the same time-series attributes as the original data, ensuring that the time-series correlation is not lost. Therefore, based on six original data sources, through the aforementioned mathematical transformations of square, cube, and square root, 18 expanded-dimensional data sources are obtained. And / or, based on feature interaction processing methods, time-series data from a second target data source and a third target data source are processed, where the second or third target data source can be any one of multiple data sources. For example, taking a feature interaction processing method that fuses two features as an example, for time-series data from any two data sources, time-series data reflecting the synergistic effects of parameters are constructed through feature interaction operations. For instance, multiplying the time-series data of rotational speed and torque yields power-related interaction features; multiplying the time-series data of current and voltage yields power consumption-related interaction features. By using feature interaction, the effectiveness of feature correlation is ensured while controlling the scale of the expanded-dimensional data sources. Therefore, based on 6 original data sources, time-series data from 15 expanded-dimensional data sources can be generated through pairwise feature interaction methods.

[0051] Correspondingly, based on the same preset processing method described above, the contribution weights of each data source can also be synchronously extended to generate the contribution weights corresponding to each expanded-dimensional data source. Specifically, if a mathematical transformation method is used to process the time series data of the first target data source, the contribution weight processing method is determined based on the mathematical transformation method, and the contribution weights of the first target data source are processed according to the contribution weight processing method to obtain the contribution weights of the expanded-dimensional data sources corresponding to the first target data source; if a feature interaction processing method is used to process the time series data of the second and third target data sources, the contribution weights of the processed expanded-dimensional data sources are determined based on the contribution weights of the second and third target data sources. For example, if the preset processing method for the first target data source (e.g., torque) is a squared mathematical transformation method, and the contribution weight of the first target data source is w... i Then, the contribution weight of the extended-dimensional data source (such as the square of torque) corresponding to the first target data source is determined by the square transformation method, which is w.i2 If the contribution weights of any two original data sources, namely the second target data source and the third target data source (e.g., torque and vehicle speed), are w respectively... i w j Then, the contribution weight of the extended-dimensional data source (i.e., the interaction between torque and vehicle speed) generated by its feature interaction can be expressed as the product of the contribution weight of torque and the contribution weight of vehicle speed (w). i · w j The above synchronous expansion process enables the synchronous adaptation of feature dimension expansion and contribution weight of time series data from various data sources. This not only mines potential correlation information in the data to enrich feature expression, but also ensures that the importance quantification of the expanded data source after processing is always consistent with the logic of the original data source, providing high-quality support for the construction of input features for subsequent temperature prediction models.

[0052] In one embodiment, the above-mentioned construction of the input features of the temperature prediction model based on the time-series data of each data source and the contribution weights of each data source, the time-series data of each extended-dimensional data source, and the contribution weights of each extended-dimensional data source includes: multiplying the time-series data of each data source with the contribution weights of each data source to obtain the feature values ​​of the dimensional features corresponding to each data source; multiplying the time-series data of each extended-dimensional data source with the contribution weights of each extended-dimensional data source to obtain the feature values ​​of the dimensional features corresponding to each extended-dimensional data source; and constructing the input features of the temperature prediction model based on the feature values ​​of the dimensional features corresponding to each data source and the feature values ​​of the dimensional features corresponding to each extended-dimensional data source.

[0053] In this embodiment, the time-series data of each data source (such as vehicle speed, torque, rotational speed, current, voltage, and ambient temperature) is multiplied by the contribution weight of the corresponding data source. The product of the time-series data of each source and the corresponding contribution weight is used as the feature value of the dimensional feature corresponding to each data source. Correspondingly, the time-series data of each expanded-dimensional data source obtained through a preset processing method (mathematical transformation and / or feature interaction operation) is multiplied by the contribution weight of the corresponding expanded-dimensional data source obtained through synchronous expansion. The product of the time-series data of each expanded-dimensional data source and the corresponding contribution weight is used as the feature value of the dimensional feature corresponding to each expanded-dimensional data source. Finally, based on the feature values ​​of the dimensional features corresponding to each data source and the feature values ​​of the dimensional features corresponding to each expanded-dimensional data source, the input features of the temperature prediction model containing multiple dimensional features are constructed.

[0054] For example, the time series data from each data source are multiplied and fused with the contribution weights of the corresponding data source, and the time series data from each expanded-dimensional data source obtained through dimensional expansion processing such as squaring, cube-cube, square root, and pairwise interactions are multiplied and fused with the contribution weights of the corresponding expanded-dimensional data source to construct the input features of the temperature prediction model. The calculation process is as follows: ; in, The feature is obtained by multiplying and fusing the time series data of any data source with the corresponding contribution weight of the data source. , , , W represents the features obtained by multiplying and fusing the square, cube, square root, and pairwise interaction weights of the contribution weights with the corresponding expanded-dimensional data sources. ij F represents the contribution weight obtained after interacting with the i-th contribution weight and the j-th contribution weight. ij The feature obtained after the interaction of the scores of the i-th feature and the j-th feature.

[0055] Therefore, the above methods can generate input features with higher representativeness and importance, which can not only enhance the influence of high-contribution data sources or high-contribution extended-dimensional data sources on the temperature prediction of the range extender, but also suppress the interference of noise information in low-contribution data sources or low-contribution extended-dimensional data sources on the model, and provide high-quality and highly focused information input for the subsequent model to accurately learn the temperature change law.

[0056] In a specific embodiment, such as Figure 5 As shown, Figure 5 This diagram illustrates a process from raw data sources to the input features of a generated temperature prediction model. First, time-series data from six data sources—vehicle speed, torque, RPM, current, voltage, and ambient temperature—are collected. Then, based on this time-series data, the Shapley additive interpretation module calculates the contribution value (SHAP value) of each time-series data point to the range extender temperature prediction for each data source. Finally, the contribution is calculated using a formula... The average absolute contribution value of each data source is calculated, and then normalized to obtain the contribution weight W of each data source. iThis process quantifies the impact of different data sources on range extender temperature prediction. Then, using pre-defined processing methods (including feature interaction operations such as squaring, cube-forming, square root, and pairwise multiplication), each data source and its contribution weight are processed to obtain the time-series data and contribution weights of each extended-dimensional data source. Finally, based on the time-series data and contribution weights of each data source, and the time-series data and contribution weights of each extended-dimensional data source, the input features H of the temperature prediction model are constructed. The final input feature H has a size of (39, 1000), where 39 represents the number of feature dimensions in H, and 1000 represents the number of time-series data samples corresponding to each dimension. This processed data can be used to construct the input features of the subsequent temperature prediction model, allowing the model to more fully extract temperature-related information from the data. It is understood that this application embodiment only uses time-series data from six data sources—vehicle speed, torque, rotational speed, current, voltage, and ambient temperature—as an example of multi-source data for illustration. In other embodiments, the collected multi-source data may include more or fewer data sources, and this application embodiment is not limited in this regard.

[0057] In one embodiment, before step S205 above, that is, before determining the positional encoding weights of the feature values ​​of each dimension feature in the input features, a range extender temperature prediction method further includes: dividing the input features into time windows to obtain sub-input features for each time window, each sub-input feature containing feature values ​​corresponding to multiple consecutive time points in each time window; determining the positional encoding weights of the feature values ​​of each dimension feature in the input features includes: for any sub-input feature, performing positional encoding calculations on the feature values ​​of each dimension feature based on the parity of the index of each dimension feature, the feature values ​​of each dimension feature, and the positional index of the feature values ​​of each dimension feature to obtain the positional encoding weights of the feature values ​​of each dimension feature.

[0058] In this embodiment, to enable the temperature prediction model to better understand the impact of time-series data of input features of various dimensions on temperature prediction, before executing step S205 (determining the positional encoding weights of feature values ​​of each dimension in the input features), the input features need to be divided into time windows. Then, positional encoding weights are calculated based on the divided sub-input features, assigning a unique vector to the position of each feature value. This vector, along with the actual content at that position, is passed to the model, allowing the model to understand the order of the input sequence. Specifically, for the input features of the temperature prediction model, time windows are divided according to the time length, decomposing the input features into multiple sub-input features containing feature values ​​of input features at multiple consecutive time points. For example, if the time series length of the original input features is 1000 (i.e., containing feature values ​​of 1000 consecutive time points), and each time window contains 100 consecutive time points, it can be evenly divided into 10 time windows, corresponding to the generation of 10 sub-input features, each with dimensions (39, 100).

[0059] Understandably, the window size can be flexibly adjusted based on the frequency of changes in the range extender's operating conditions. If the range extender's operating conditions fluctuate frequently (such as frequent starts and stops, sudden load changes), a smaller window size (such as 50 time points) can be set to capture the impact of short-term operating condition changes on temperature more precisely; if the operating conditions are relatively stable (such as smooth operation during constant speed driving), the window size can be appropriately increased (such as 200 time points) to reduce the computational load while ensuring information integrity.

[0060] Furthermore, for each sub-input feature, based on the parity of the index of each dimension feature, the feature value of each dimension feature, and the position index of the feature value of each dimension feature, position encoding calculation is performed on the feature value of each dimension feature to obtain the position encoding weight of the feature value of each dimension feature. This quantifies the position information of different dimension features within the time window and its correlation with the feature value, providing a basis for the importance of position dimension for subsequent adaptive feature fusion. Specifically, firstly, the key parameters in each sub-input feature are defined: dimension feature index (i.e., the index of each dimension in the sub-input feature, such as index 1 to 39 for 39-dimensional features), feature value position index (i.e., the index of each time point within the time window, such as index 0 to 99 for 100 time points), and feature value of each dimension feature at the corresponding position index; secondly, an appropriate encoding function is selected according to the parity of the dimension feature index (e.g., a sine function is used when the index is even, and a cosine function is used when the index is odd), the encoding cardinality is adjusted in combination with the feature value, and the encoding frequency is controlled by the feature value position index, finally generating the position encoding weight of each dimension feature at each position index. For example, for any sub-input feature, the positional encoding of the feature value under each feature dimension is calculated as follows: , ; in, , The location encoding weights represent the feature values ​​at different positions in different feature dimensions. `pos` represents the index of each feature dimension, `j` represents the index of the j-th feature value, `Lxj` represents the j-th feature value, and `dmodel=100` represents the length of the time series. The value of `j` is 0 - `dmodel / 2`. If the dimension `pos` is even, a sine function is used; if the dimension `pos` is odd, a cosine function is used. Using sine and cosine functions generates periodic variations with different frequencies, thus providing a unique code for each position. The sine and cosine functions have different frequencies in different dimensions (i.e., different `pos`), resulting in different patterns in the location encoding across different dimensions. This helps the model distinguish different location information. Through these calculations, the encoding of the feature value at each feature dimension `pos` is obtained. The final calculated location encoding size is (10, 39, 100), where 10 represents 10 time windows, 39 represents 39 feature dimensions, and 100 is the number of location encoding value weights. The above-mentioned method distinguishes the positional attributes of features in different dimensions by the periodic differences of sine and cosine functions, and dynamically adjusts the encoding results by combining feature values. This allows the positional encoding weights to reflect both the temporal and positional relationships of features and the magnitude of feature values, laying the foundation for subsequent adaptive feature fusion based on positional information and helping the model to more accurately capture the spatiotemporal patterns of temperature changes in the range extender.

[0061] In one embodiment, such as Figure 6 As shown, the adaptive fusion processing of the input features based on the positional encoding weights of the feature values ​​of each dimension, as described in step S205 above, to obtain the fused features, may include the following steps: S601 calculates the similarity between any two dimensional features based on the positional encoding weights of any two dimensional features.

[0062] S602, construct a similarity matrix of the input features based on the similarity between any two dimensional features in the input features.

[0063] S603 uses the similarity matrix as the adjacency matrix of the feature locations in the graph convolutional neural network.

[0064] S604 uses a graph convolutional neural network to perform graph convolution processing on the adjacency matrix of feature locations and the input features to obtain fused features.

[0065] In this embodiment, the similarity between any two dimensional features is first calculated based on their positional encoding weights. Specifically, the positional encoding weights of each dimensional feature are first considered as the positional encoding vector of that dimensional feature. For example, taking a sub-input feature of any time window as an example, the positional encoding vector of each dimensional feature in the sub-input feature has the same dimension as the length of the time window of the sub-input feature (e.g., if the time window contains 100 time points, then the positional encoding vector is 100-dimensional), and each element in the vector corresponds to the positional encoding weight of that dimensional feature at a certain time point. For example, for the dimensional feature "torque squared" in the sub-input feature, its positional encoding weights at 100 time points are PE1, PE2, ..., PE 100 Then the position encoding vector of this dimension feature is [PE1, PE2, ..., PE... 100 Furthermore, after obtaining the positional encoding vectors of each dimension feature, a similarity matrix for the sub-input feature is constructed by calculating the similarity (e.g., cosine similarity) between any two dimension feature positional encoding vectors. For example, cosine similarity is used as the similarity calculation index (cosine similarity ranges from -1 to 1; the closer the value is to 1, the more consistent the directions of the two vectors are, i.e., the stronger the positional correlation of the corresponding dimension features), and cosine similarity is calculated for any two dimension feature positional encoding vectors. If the sub-input feature contains 39 dimension features, the final constructed similarity matrix is ​​a 39×39 matrix, where the element (i,j) represents the similarity between the positional encoding vectors of the i-th dimension feature and the j-th dimension feature.

[0066] Then, the similarity matrix is ​​used as the feature location adjacency matrix of the graph convolutional neural network (GCN). The GCN then performs graph convolution processing on the feature location adjacency matrix and the input features to obtain the fused features. Specifically, the feature values ​​of the sub-input features of each time window are used as the node features of each graph node in the graph structure, and the feature location adjacency matrix of the corresponding time window is used as the adjacency matrix of the graph structure to define the connection relationships and strengths between graph nodes. The GCN then performs neighborhood information aggregation and nonlinear transformation on the node features based on the adjacency matrix to generate the sub-fused features of each time window, which together constitute the fused features of the input features of the input temperature prediction model. For example, the calculation formula for adaptive fusion processing using the GCN is as follows: ; in, For sub-input features, This is the corresponding weight matrix. As a sub-fusion feature, A fL The adjacency matrix of the characteristic positions, For the adjacency matrix A of the characteristic positions fL A symmetric normalization operation is performed, and ReLU is the activation function.

[0067] In the Graph Convolutional Neural Network (GCN) process, firstly, the adjacency matrix AfL of the feature locations is symmetrically normalized to eliminate the interference of degree differences between nodes (dimensional features) on information aggregation. Then, the normalized adjacency matrix is ​​multiplied by the node feature matrix of the sub-input feature and the GCN weight matrix W1 (the weight matrix W1 is determined according to the model design, e.g., 100×50) to achieve linear transformation of node features and fusion of neighborhood information. After linear transformation and neighborhood aggregation, the ReLU nonlinear activation function is introduced to process the fused features, ultimately generating sub-fused features. The dimension of these sub-features is determined by the output dimension of the GCN. For example, if the output dimension of the GCN is 50, then the size of the output sub-fused feature H2 is (39, 50). Therefore, this application embodiment constructs a feature location adjacency matrix by calculating the similarity between features, which can accurately measure the correlation of 39-dimensional features in the location encoding space. Combined with a graph convolutional neural network, the sub-input features of each time window are adaptively fused based on the similarity information in the adjacency matrix. The graph convolutional network can propagate and update information based on feature connection strength, realize the weighted aggregation of related features and the suppression of irrelevant features, effectively optimize the model's expressive power and feature selection, and significantly improve the accuracy of range extender temperature prediction.

[0068] For example, such as Figure 7The diagram illustrates a specific process flow for adaptive fusion according to an embodiment of this application. First, the input feature H with a feature size of (39, 1000) is divided into 10 time windows (e.g., time window 1, time window 2, etc.). The feature size of the sub-input feature in each time window is (39, 100), meaning each window contains 39 dimensional features, and each dimension has 100 consecutive time point feature values. Then, for the sub-input feature of each time window, a positional encoding formula is used... , First, calculate the positional weights of the feature values ​​for each dimension to form an equal-positional encoding matrix. Then, based on the positional encoding of each dimension, calculate the cosine similarity between the positional encoding vectors of any two dimensions to construct the feature positional adjacency matrix for each time window, such as A. fL1 A fL2 ,…, A fL10 (Each feature location adjacency matrix is ​​39×39 in size) to characterize the positional correlation between features. Finally, the feature values ​​of the sub-input features of each time window are used as the graph node features of each graph node in the graph structure, and the adjacency matrix of the feature locations corresponding to the time window is used as the adjacency matrix of the graph structure. These are input into a graph convolutional neural network (GCN) (such as GCN1, GCN2, etc.). The graph convolutional neural network performs neighborhood information aggregation and nonlinear transformation on the node features based on the adjacency matrix to generate sub-fused features for each time window (such as H). 2_1 H 2_1 ,…, H 2_10 This enables the effective aggregation of relevant features and the suppression of irrelevant features, providing optimized feature inputs for subsequent range extender temperature prediction.

[0069] In one embodiment, such as Figure 8 As shown, prior to step S206 above, a range extender temperature prediction method further includes the following steps: S801, for any time window of sub-fusion features, determine the contribution weight of each dimension of the sub-fusion features to the range extender temperature prediction.

[0070] S802, construct a contribution matrix based on the contribution weights of each dimension feature in the sub-fusion features to the range extender temperature prediction.

[0071] S803 calculates the similarity between any two dimensional features in the sub-fusion features and constructs a feature similarity matrix based on the similarity between any two dimensional features in the sub-fusion features.

[0072] S804 performs collaborative filtering based on the contribution matrix and feature similarity matrix to select multiple target dimensional features whose dimensional feature correlation is greater than a set threshold.

[0073] S805 constructs fused features based on the feature values ​​of each target dimension.

[0074] In this embodiment, to further optimize the quality of input features and remove redundant dimensional features, before executing step S206, i.e., before inputting the fused features into the temperature prediction model, the sub-fused features of each time window can be filtered to obtain the fused features after removing redundant dimensional features. Specifically, firstly, for the sub-fused features of any time window, such as... Figure 9 As shown, the sub-fusion feature H of size (39, 50) is obtained after adaptive fusion processing by a graph convolutional neural network. 2_i The contribution weights of each dimension of the sub-fusion features to the range extender temperature prediction are determined using feature importance assessment methods such as SHAP (Shapley Additive Interpretation). The step of determining the contribution weights of each dimension of the sub-fusion features to the range extender temperature prediction using SHAP is similar to step S302 above; specific steps can be found in the relevant description of step S302 above, and will not be repeated here. The contribution weights of each feature dimension to the range extender temperature prediction calculated by the above method can be 39 contribution weights with a value between (0, 1), denoted by w. H2_i Indicate. Then w H2_i Multiplying it by its corresponding transpose matrix yields the contribution matrix A. c The calculation process is as follows:

[0075] ; in, for The contribution transpose matrix, A c A contribution matrix of size 39×39 is generated. Simultaneously, a feature similarity matrix is ​​constructed by calculating the similarity (e.g., cosine similarity) between any two dimensional features in the sub-fusion feature. If the sub-fusion feature has 39 dimensional features, then the constructed feature similarity matrix A... S This is a matrix of size 39×39. The elements S in the matrix... m,n This represents the similarity between the m-th and n-th dimension features. The closer the value is to 1, the stronger the linear correlation between the two dimension features.

[0076] Furthermore, the contribution matrix A c and feature similarity matrix A S Perform a product interaction operation to obtain a product interaction matrix. For example, combine a 39×39 contribution matrix Ac with a 39×39 feature similarity matrix A. SElement-wise multiplication yields a 39×39 product interaction matrix. This matrix is ​​then binarized. Specifically, binarization involves setting a threshold τ (the optimal value can be determined through validation set experiments, e.g., τ=0), setting elements with values ​​greater than τ to 1 and those less than τ to 0. Then, based on the binarized matrix, the number of non-zero elements (i.e., elements with a matrix value of 1) in each row is counted. This count represents the corresponding dimension feature count in the sub-fusion feature set. A higher count indicates a strong correlation between this dimension feature and other dimensions, considering feature importance and similarity; conversely, a lower count indicates a lower correlation and potential redundancy.

[0077] Furthermore, based on the count values, the dimensional features in the sub-fusion features are filtered to identify multiple target dimensional features whose correlation is greater than a set threshold. Then, based on the feature values ​​of each target dimensional feature, a fusion feature after removing redundant dimensional features is constructed. Specifically, the count values ​​of each dimensional feature are sorted in descending order. The dimensional features with the top_K count values ​​(i.e., the top K count values, where K is a positive integer) are considered valid features and retained. For example, the top 20 count values ​​are retained to filter redundant dimensional features, resulting in multiple target dimensional features whose correlation is greater than a set threshold. The fusion feature after removing redundant dimensional features is then constructed based on the feature values ​​of each target dimensional feature. Through these steps, redundant dimensions caused by high correlation in the sub-fusion features can be effectively identified and removed. While retaining core effective information, the feature dimensionality and complexity are reduced, providing simpler and more representative input features for the subsequent range extender temperature prediction model, thus helping to improve model training efficiency and prediction accuracy.

[0078] In one embodiment, such as Figure 10 As shown, after step S805 above, i.e., constructing fused features based on the feature values ​​of each target dimension, a range extender temperature prediction method may further include the following steps: S1001, for the fusion features of each time window, determine the position encoding weight of each time window based on the count value of each dimension feature in the fusion features.

[0079] S1002, global temporal information fusion processing is performed on the fusion features of each time window according to the position encoding weight of each time window to obtain the target fusion features.

[0080] In this embodiment, after step S805 (constructing fusion features based on feature values ​​of each target dimension), to further integrate the fusion features of each time window and explore global temporal patterns, global temporal information fusion can be achieved through the following steps to obtain target fusion features. Specifically, for the fusion feature of any time window (such as a filtered fusion feature of size (20,50) after removing redundant dimension features), firstly, the count values ​​of each dimension feature in the fusion feature are obtained and normalized to obtain the weights of each dimension feature in the fusion feature. Further, based on the parity of each time window index, the weights of each dimension feature in the filtered fusion feature of each time window, and the position index of each dimension feature, position encoding calculation is performed on each time window to obtain the position encoding weight of each time window. The calculation formula is as follows: , ; in, The positional encoding weights represent different time windows, where T represents the index of the time window, k is the positional index of the k-th dimension feature, and the value of k ranges from 0 to dT / 2. xk d represents the weights of each dimension of the fused features within the corresponding time window. T =20 represents 20 weights. Specifically, the location encoding calculation logic for the time window is similar to the location encoding of the dimensional features in step S205. For example, if the index T of the time window is even, the location encoding weight of the time window is calculated using the sine function encoding formula; if the index T of the time window is odd, the location encoding weight of the time window is calculated using the cosine function encoding formula. Through this encoding method, both parity is used to distinguish the temporal attributes of different time windows, and the weight information of each dimensional feature is incorporated, so that the calculated location encoding weight can reflect the dual correlation between the temporal sequence and dimensional importance of the time window.

[0081] Furthermore, based on the positional encoding weights of each time window, global temporal information fusion processing is performed on the filtered fusion features of all time windows to obtain the target fusion features. Specifically, firstly, the cosine similarity between the positional encoding weights of each time window is calculated, thereby obtaining a 10×10 temporal adjacency matrix, which is then used as A. T This is used to quantify the correlation strength of different time windows in the location encoding space. Then, the temporal location adjacency matrix A is... T The filtered and fused features from each time window are input into the graph convolutional network to efficiently fuse global temporal information. The specific calculation process is as follows: ; Among them, H outHere, W2 represents the target fusion feature, W3 represents the corresponding weight matrix, and H3 represents the fusion feature. For the time position adjacency matrix A T A symmetric normalization operation is performed, and ReLU is the activation function.

[0082] In the process of graph convolutional neural network (GCN) processing, the temporal adjacency matrix A is... T Perform symmetric normalization operation (i.e.) To eliminate the interference of different time window degree differences on temporal information aggregation, the normalized temporal adjacency matrix is ​​multiplied with the fusion feature matrix H3 and the GCN weight matrix W2 (the weight matrix W2 is determined according to the model design, for example, 20×25) to achieve linear transformation of temporal features and cross-time window information fusion. After linear transformation and global temporal aggregation, the ReLU nonlinear activation function is introduced to process the fused features, finally generating the target fusion feature H. out Its dimension is determined by the set output dimension of the GCN. For example, if the output dimension of the GCN mentioned above is 25, then the output target fusion feature H out The size is (10, 20, 25).

[0083] Therefore, the embodiments of this application can adaptively perform global temporal fusion of the fusion features of each time window based on the positional encoding weights of each time window. This achieves weighted aggregation of relevant temporal features and suppression of irrelevant temporal features, effectively reducing the loss of important information and timely capturing abnormal changes at critical moments. It optimizes the model's performance in the feature extraction stage, enhances the model's ability to understand and infer complex relationships in time-series data, and significantly improves the accuracy, reliability, and robustness of range extender temperature prediction. Furthermore, by leveraging feature dimensionality expansion and collaborative filtering mechanisms, redundant features can be effectively filtered out, improving computational efficiency and avoiding overfitting.

[0084] Further, in one embodiment, step S206 above, which involves inputting the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model, includes: inputting the target fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0085] In this embodiment, the target fusion feature, optimized through the entire process of original data acquisition, contribution weight calculation, data dimensionality expansion, time window division, position encoding, graph convolution adaptive fusion, redundant feature filtering, and global temporal fusion, is input into the pre-trained temperature prediction model. For example, the target fusion feature H3 with a feature size of (10, 20, 25) (corresponding to 10 time windows, 20 core dimensional features per window, and 50-dimensional information processed by graph convolution for each feature) is used as the input feature of the temperature prediction model. This target fusion feature H3 integrates key information, temporal correlation patterns, and feature importance weights from the multi-source data of the range extender, effectively eliminating redundant interference and possessing high information density and strong temperature representation capabilities. Furthermore, the temperature prediction model receives feature information through the input layer, performs nonlinear transformation and information integration through the hidden layer (e.g., 3 layers, 256 neurons per layer), and gradually explores the mapping relationship between the target fusion feature H3 and the range extender temperature. Finally, the predicted temperature is output through the output layer of the temperature prediction model, completing the result output. It should be noted that, in this embodiment, the predicted temperature output by the model can be either the predicted temperature value for the current time period or the predicted temperature value for the next time period, depending on the model training objectives and the requirements of the actual application scenario. This embodiment does not limit this.

[0086] Furthermore, after obtaining the predicted temperature from the temperature prediction model, a temperature warning can be issued based on the predicted temperature and the preset temperature threshold of the range extender. If the predicted temperature is higher than the range extender's temperature threshold (e.g., 120℃, which can be set according to the range extender's hardware high-temperature resistance limit), a temperature warning is triggered, guiding operators to take cooling measures (e.g., reducing load, activating the cooling system). If the predicted temperature is lower than the range extender's temperature threshold, the predicted temperature is synchronized to the range extender temperature monitoring platform, an important indicator for evaluating the range extender's operating status, and assists in optimizing the range extender's operating parameters and maintenance strategies.

[0087] It should be understood that although the steps in the flowchart are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0088] This application also provides a range extender temperature prediction device. For example... Figure 11As shown, a range extender temperature prediction device includes a data acquisition module 1101, a processing module 1102, a first determination module 1103, a first construction module 1104, a first fusion module 1105, and a prediction module 1106. The data acquisition module 1101 is used to acquire time-series data from multiple data sources of the range extender and determine the contribution weight of each data source to the range extender temperature prediction. The processing module 1102 is used to process the time-series data from each data source based on a preset processing method to obtain processed time-series data from each data source, and uses the processed time-series data from each data source as time-series data for extended-dimensional data sources. The first determination module 1103 is used to determine the contribution weight of each extended-dimensional data source based on the preset processing method and the contribution weight of each data source. The first construction module 1104 is used to construct the data based on the time-series data from each data source... The input features of the temperature prediction model are constructed by combining the contribution weights of each data source, the time series data of each extended-dimensional data source, and the contribution weights of each extended-dimensional data source. The input features include multiple dimensional features, each corresponding to a data source or an extended-dimensional data source. The first fusion module 1105 is used to determine the positional encoding weights of the feature values ​​of each dimensional feature in the input features, and performs adaptive fusion processing on the input features according to the positional encoding weights of the feature values ​​of each dimensional feature to obtain fused features. The prediction module 1106 is used to input the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0089] In one embodiment, the acquisition module 1101 is specifically used to: acquire sample time-series data from each data source and sample temperature data of the range extender corresponding to the sample time-series data from each data source; and obtain the contribution weight of each data source based on the sample time-series data from each data source, the sample temperature data of the corresponding range extender, the temperature prediction model, and using Shapley additive interpretation.

[0090] In one embodiment, the preset processing method includes a set mathematical transformation method and / or a feature interaction processing method. The processing module 1102 is specifically used to: process the time series data of a first target data source based on the mathematical transformation method, wherein the first target data source is one or more; and / or, process the time series data of a second target data source and a third target data source based on the feature interaction processing method, wherein the second target data source or the third target data source is any one of multiple data sources; the first determining module 1103 is specifically used to: if the mathematical transformation method is used to process the time series data of the first target data source, then determine the contribution weight processing method based on the mathematical transformation method, and process the contribution weight of the first target data source according to the contribution weight processing method to obtain the contribution weight of the extended dimension data source corresponding to the first target data source; if the feature interaction processing method is used to process the time series data of the second target data source and the third target data source, then determine the contribution weight of the processed extended dimension data source based on the contribution weight of the second target data source and the contribution weight of the third target data source.

[0091] In one embodiment, the first construction module 1104 is specifically used to: multiply the time series data of each data source with the contribution weight of each data source to obtain the feature value of the dimensional feature corresponding to each data source; multiply the time series data of each extended dimension data source with the contribution weight of each extended dimension data source to obtain the feature value of the dimensional feature corresponding to each extended dimension data source; and construct the input features of the temperature prediction model based on the feature values ​​of the dimensional features corresponding to each data source and the feature values ​​of the dimensional features corresponding to each extended dimension data source.

[0092] In one embodiment, a range extender temperature prediction device further includes: a partitioning module, used to partition the input features according to time windows to obtain sub-input features for each time window, each sub-input feature containing feature values ​​corresponding to multiple consecutive time points in each time window; and a first fusion module 1105, specifically used to: for any sub-input feature, perform position encoding calculation on the feature values ​​of each dimension feature based on the parity of the feature index of each dimension, the feature value of each dimension feature, and the position index of the feature value of each dimension feature, to obtain the position encoding weight of the feature value of each dimension feature.

[0093] In one embodiment, the first fusion module 1105 is specifically used to: calculate the similarity between any two dimensional features based on the positional encoding weights of any two dimensional features; construct a similarity matrix of the input features based on the similarity between any two dimensional features in the input features; use the similarity matrix as the feature position adjacency matrix of the graph convolutional neural network; and perform graph convolution processing on the feature position adjacency matrix and the input features through the graph convolutional neural network to obtain the fused features.

[0094] In one embodiment, a range extender temperature prediction device further includes: a second determining module, configured to determine the contribution weight of each dimension feature in the sub-fusion features for range extender temperature prediction for any given time window; a second constructing module, configured to construct a contribution matrix based on the contribution weight of each dimension feature in the sub-fusion features for range extender temperature prediction; a third constructing module, configured to calculate the similarity between any two dimension features in the sub-fusion features and construct a feature similarity matrix based on the similarity between any two dimension features in the sub-fusion features; a collaborative filtering module, configured to perform collaborative filtering based on the contribution matrix and the feature similarity matrix to select multiple target dimension features whose dimensional feature correlation is greater than a set threshold; and a fourth constructing module, configured to construct fusion features based on the feature values ​​of each target dimension feature.

[0095] In one embodiment, a range extender temperature prediction device further includes: a third determining module, used to determine the position encoding weight of each time window based on the count value of each dimension feature in the fusion features for each time window; a second fusion module, used to perform global time-series information fusion processing on the fusion features of each time window based on the position encoding weight of each time window to obtain target fusion features; and a prediction module 1106, specifically used to: input the target fusion features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

[0096] For specific limitations regarding the temperature prediction device for a range extender, please refer to the limitations of the temperature prediction method for a range extender mentioned above, which will not be repeated here. The modules in the aforementioned user-action-based liveness detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0097] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 12 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a range extender temperature prediction method.

[0098] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device on which the present application is intended to be applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0099] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a range extender temperature prediction method.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0101] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting the temperature of a range extender, characterized in that, The method includes: Time-series data from multiple data sources for the range extender were collected, and the contribution weight of each data source to the temperature prediction of the range extender was determined. The time series data from each data source is processed based on a preset processing method to obtain the processed time series data from each data source. The processed time series data from each data source is then used as the time series data from the extended dimension data source. Based on the preset processing method and the contribution weight of each data source, the contribution weight of each extended dimension data source is determined. The input features of the temperature prediction model are constructed based on the time series data of each data source and the contribution weight of each data source, as well as the time series data of each extended dimension data source and the contribution weight of each extended dimension data source. The input features include multiple dimensional features, each dimensional feature corresponding to each data source or each extended dimension data source. Determine the positional encoding weights of the feature values ​​of each dimension of the input features, and perform adaptive fusion processing on the input features based on the positional encoding weights of the feature values ​​of each dimension to obtain fused features; The fused features are input into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

2. The method according to claim 1, characterized in that, The determination of the contribution weight of each data source to the range extender temperature prediction includes: Collect sample time-series data from each data source and sample temperature data of the range extender corresponding to the sample time-series data from each data source; Based on the sample time-series data from each data source, the sample temperature data of the corresponding range extender, and the temperature prediction model, and using Shapley additive interpretation, the contribution weights of each data source are obtained.

3. The method according to claim 1, characterized in that, The preset processing method includes a set mathematical transformation method and / or feature interaction processing method. The processing of time-series data from each data source based on the preset processing method includes: The time-series data of the first target data source is processed based on the mathematical transformation method, wherein the first target data source is one or more. And / or, based on the feature interaction processing method, the time-series data of the second target data source and the third target data source are processed, wherein the second target data source or the third target data source is any one of the plurality of data sources; The determination of the contribution weight of each expanded-dimensional data source based on the preset processing method and the contribution weight of each data source includes: If the mathematical transformation method is used to process the time series data of the first target data source, then the contribution weight processing method is determined based on the mathematical transformation method, and the contribution weight of the first target data source is processed according to the contribution weight processing method to obtain the contribution weight of the extended data source corresponding to the first target data source. If the feature interaction processing method is used to process the time series data of the second target data source and the third target data source, the contribution weight of the processed extended-dimensional data source is determined based on the contribution weight of the second target data source and the contribution weight of the third target data source.

4. The method according to claim 1, characterized in that, The input features for constructing the temperature prediction model based on time-series data from each data source and the contribution weights of each data source, as well as time-series data from each extended-dimensional data source and the contribution weights of each extended-dimensional data source, include: The time series data of each data source is multiplied by the contribution weight of each data source to obtain the feature value of the dimensional feature corresponding to each data source. Multiply the time series data of each extended dimension data source with the contribution weight of each extended dimension data source to obtain the feature value of the dimensional feature corresponding to each extended dimension data source. The input features of the temperature prediction model are constructed based on the feature values ​​of the dimensional features corresponding to each data source and the feature values ​​of the dimensional features corresponding to each extended-dimensional data source.

5. The method according to claim 1, characterized in that, Before determining the positional encoding weights of the feature values ​​of each dimension of the input features, the method further includes: The input features are divided into time windows to obtain sub-input features for each time window. Each sub-input feature contains feature values ​​corresponding to multiple consecutive time points in each time window. The determination of the positional encoding weights of the feature values ​​of each dimension of the input features includes: For any sub-input feature, based on the parity of the index of each dimension feature, the feature value of each dimension feature, and the position index of the feature value of each dimension feature, the position encoding weight of the feature value of each dimension feature is calculated.

6. The method according to claim 5, characterized in that, The adaptive fusion processing of the input features based on the positional encoding weights of the feature values ​​of each dimension to obtain fused features includes: The similarity between any two dimensional features is calculated based on the positional encoding weights of any two dimensional features; Construct a similarity matrix for the input features based on the similarity between any two dimensional features in the input features; The similarity matrix is ​​used as the feature location adjacency matrix of the graph convolutional neural network; The fused features are obtained by performing graph convolution processing on the adjacency matrix of the feature positions and the input features using the graph convolutional neural network.

7. The method according to claim 6, characterized in that, Before inputting the fused features into the temperature prediction model, the method further includes: For any given time window, determine the contribution weight of each dimension of the sub-fusion features to the range extender temperature prediction. Contribution matrix is ​​constructed based on the contribution weights of each dimension feature in the sub-fusion features to the range extender temperature prediction; Calculate the similarity between any two dimensional features in the sub-fusion feature, and construct a feature similarity matrix based on the similarity between any two dimensional features in the sub-fusion feature; Collaborative filtering is performed based on the contribution matrix and the feature similarity matrix to select multiple target dimensional features whose dimensional feature relevance is greater than a set threshold; The fused features are constructed based on the feature values ​​of each target dimension.

8. The method according to claim 7, characterized in that, After constructing the fused feature based on the feature values ​​of each target dimension, the method further includes: For the fusion features of each time window, the position encoding weight of each time window is determined based on the count value of each dimension feature in the fusion features; Based on the positional encoding weights of each time window, global temporal information fusion processing is performed on the fusion features of each time window to obtain the target fusion features; The step of inputting the fused features into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model includes: The target fusion features are input into the temperature prediction model to obtain the predicted temperature of the range extender output by the temperature prediction model.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images