Neural network driven electric vehicle temperature control strategy optimization method and system

Abstract

The application provides a neural network driven electric vehicle temperature control strategy optimization method and system. By performing feature representation on a first number of candidate electric vehicle battery temperature control strategy corresponding first number of strategy trigger monitoring sequences, a target feature vector corresponding to each strategy trigger monitoring sequence is obtained; a second number of target hybrid feature vectors are determined according to a trigger prediction reference value and loaded into a hybrid trigger prediction network, and a second number of to-be-determined hybrid trigger prediction probabilities are output based on the hybrid trigger prediction network; a third number of target hybrid trigger prediction probabilities are determined according to a hybrid trigger prediction reference value, and each target hybrid trigger prediction probability is weighted according to a pre-determined weighting coefficient, and the candidate electric vehicle battery temperature control strategy corresponding to the maximum target hybrid trigger prediction probability after weighting is taken as the electric vehicle temperature control strategy. The application can improve the scientificity of electric vehicle battery temperature control, and is beneficial to battery health maintenance and life improvement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing and machine learning, and particularly relates to a neural network driven electric vehicle temperature control strategy optimization method and system. BACKGROUND

[0002] With the rapid expansion and wide application of the electric vehicle market, the performance and stability of the electric vehicle have attracted widespread attention, and the temperature control strategy plays a crucial role in the performance, life and safety of the electric vehicle battery. When operating in different environmental temperatures, the temperature control of the battery faces great challenges. In a high-temperature environment, the battery is prone to overheating, which accelerates the chemical reaction inside the battery, leading to rapid capacity attenuation, increased internal resistance, reduced charging and discharging efficiency of the battery, and even safety problems such as thermal runaway. In a low-temperature environment, the activity of the battery is reduced, the conductivity of the electrolyte is poor, and the output power and capacity of the battery are greatly reduced, which significantly shortens the cruising range of the electric vehicle, and even causes the electric vehicle to fail to start. The existing electric vehicle temperature control strategy lacks accurate prediction and optimization selection capability for the triggering probability of different types of temperature control strategies. In actual application, there are various temperature control strategies to choose from, such as active cooling, passive cooling, active heating, etc., and each strategy contains different execution types. The current technology cannot accurately determine which temperature control strategy should be selected under which conditions, resulting in low temperature control efficiency and failing to fully utilize the advantages of various temperature control strategies. SUMMARY

[0003] The present application provides a neural network driven electric vehicle temperature control strategy optimization method and system.

[0004] According to an aspect of the present application, a neural network driven electric vehicle temperature control strategy optimization method is provided, comprising: performing feature representation on a first number of candidate electric vehicle battery temperature control strategy corresponding to a first number of strategy trigger monitoring sequences respectively, to obtain a target feature vector corresponding to each of the strategy trigger monitoring sequences; determining a second number of target hybrid feature vectors from among the first number of target feature vectors according to a trigger prediction reference value, the second number being a positive integer not greater than the first number; loading the second number of target hybrid feature vectors into a hybrid trigger prediction network respectively, and outputting a second number of pending hybrid trigger prediction probabilities based on the hybrid trigger prediction network; determining a third number of target hybrid trigger prediction probabilities from among the second number of pending hybrid trigger prediction probabilities according to a hybrid trigger prediction reference value, wherein the third number is a positive integer not greater than the second number; and weighting the third number of target hybrid trigger prediction probabilities according to the respective weighting coefficients of the candidate electric vehicle battery temperature control strategies pre-determined for the third number of target hybrid trigger prediction probabilities, and taking the candidate electric vehicle battery temperature control strategy corresponding to the maximum target hybrid trigger prediction probability after weighting as the electric vehicle temperature control strategy.

[0005] According to another aspect of the present application, there is provided a temperature control strategy optimization system, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0006] The present application has at least the following beneficial effects: in the network training process, the present application is based on the feature representation of the first number of policy trigger monitoring sequence samples to obtain the sample feature vector corresponding to each policy trigger monitoring sequence sample, then determines the second number of mixed sample feature vectors in the first number of sample feature vectors according to the trigger prediction reference value, then loads the second number of mixed sample feature vectors into the mixed trigger prediction network respectively, outputs the second number of mixed trigger prediction probabilities based on the mixed trigger prediction network, in addition, loads the second number of mixed sample feature vectors into the collaborative trigger prediction network respectively, outputs the second number of collaborative trigger prediction probabilities based on the collaborative trigger prediction network, then obtains the first supervision cost according to the uniformity supervision requirement, and adjusts the network parameters of the mixed trigger prediction network according to the first supervision cost. Based on this, the second number of mixed sample feature vectors can be selected as the input of the network training of the mixed trigger prediction network and the collaborative trigger prediction network respectively, which can alleviate the calculation cost of the mixed trigger prediction network and improve the efficiency. In addition, according to the uniformity supervision requirement, the mixed trigger prediction network that needs to be trained can model the collaborative trigger prediction probability of each strategy type that balances the probability value of each strategy type and meets the performance requirements of the electric vehicle. On the one hand, the network parameters of the uniformity supervision can be automatically optimized in the network training process according to the first supervision cost, and on the other hand, the accuracy of the mixed trigger prediction network in predicting the trigger probability can be improved based on different types of strategies, so as to obtain a mixed trigger prediction network that meets the performance requirements of the electric vehicle, and the temperature control of the battery is more scientific, which is conducive to the health maintenance and life extension of the battery. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 A flowchart of a neural network driven electric vehicle temperature control strategy optimization method according to an embodiment of the present application is shown.

[0008] Figure 2 A composition diagram of a temperature control strategy optimization system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0009] The neural network driven electric vehicle temperature control strategy optimization method provided by the application can be executed by a temperature control strategy optimization system of an electric vehicle, such as an electronic control unit (ECU), a battery management system (BMS), and the like.

[0010] Please refer to Figure 1 The neural network driven electric vehicle temperature control strategy optimization method provided by the application includes the following steps: step S101: performing feature representation on a first number of candidate electric vehicle battery temperature control strategy corresponding first number of strategy trigger monitoring sequences respectively, to obtain a target feature vector corresponding to each strategy trigger monitoring sequence.

[0011] The candidate electric vehicle battery temperature control strategy is a series of schemes formulated for adjusting the temperature of the electric vehicle battery, including different types of strategies such as temperature increase and temperature decrease, and each type of strategy can be further divided into different execution types. For example, the temperature decrease strategy can be divided into active cooling and passive cooling, and the active cooling can be achieved by starting the refrigeration equipment, while the passive cooling can rely on natural heat dissipation. The first number represents the total number of candidate electric vehicle battery temperature control strategies.

[0012] The strategy trigger monitoring sequence is a detailed record of the triggering condition of each candidate temperature control strategy, which contains a series of information related to the strategy trigger, such as trigger time, battery temperature at the time of trigger, environment temperature, battery capacity, etc. Taking a temperature decrease strategy as an example, its strategy trigger monitoring sequence may record the specific time when the active cooling strategy is triggered and the environment temperature at that time when the battery temperature exceeds 30 degrees Celsius every day in the past week.

[0013] Feature representation is the process of converting the strategy trigger monitoring sequence into a target feature vector. The temperature control strategy optimization system needs to extract key features from the strategy trigger monitoring sequence and combine these features into a vector. When performing feature representation, the temperature control strategy optimization system can use a variety of technical means.

[0014] One of the feasible technical means is One-Hot Encoding. For categorical features, such as the execution type of the strategy (active cooling, passive cooling), One-Hot Encoding can convert each category into a binary vector. Assuming there are two execution types, active cooling is represented by the vector [1, 0], and passive cooling is represented by the vector [0, 1]. Another technical means is normalization. For numerical features, such as battery temperature, environment temperature, battery capacity, etc., since the value range of these features may differ greatly, in order to avoid the influence of certain features on the model being too large, normalization processing can be performed, such as Min-Max Normalization and Z-score normalization.

[0015] The temperature control strategy optimization system can also employ feature selection techniques to select the features that have the most influence on the strategy trigger prediction from a large number of features, in order to reduce the amount of calculation and improve the accuracy of the model. For example, by calculating the correlation between the features and the strategy trigger, the features with higher correlation are selected. After feature extraction and processing, the temperature control strategy optimization system combines these features into a vector, i.e., a target representation vector. Each dimension of the target representation vector corresponds to a feature, and its length depends on the number of selected features. Assuming that the battery temperature, ambient temperature, battery power, and strategy execution type are selected as the four features, after processing, the target representation vector can be a four-dimensional vector, such as [0.5, 0.3, 0.6, [1, 0]], where the first three dimensions correspond to the normalized battery temperature, ambient temperature, and battery power, respectively, and the last dimension is the strategy execution type represented by one-hot encoding.

[0016] Step S102: According to the trigger prediction reference value, determine a second number of target hybrid representation vectors from the first number of target representation vectors, wherein the second number is a positive integer not greater than the first number.

[0017] The trigger prediction reference value is a pre-set threshold value used to determine whether a candidate electric vehicle battery temperature control strategy corresponding to a target representation vector has enough possibility to be triggered. It plays a key filtering role in the entire screening process. For example, if the trigger prediction reference value is set to 0.3, it means that only when the trigger possibility of a target representation vector is greater than or equal to 0.3, the target representation vector can be retained for subsequent analysis. The first number of target representation vectors is the result obtained after the first number of candidate electric vehicle battery temperature control strategies are represented by features in step S101. Each target representation vector represents the relevant feature information of a candidate temperature control strategy. Assuming that the first number is 10, there are 10 target representation vectors, each corresponding to a different candidate electric vehicle battery temperature control strategy.

[0018] When determining the second number of target hybrid representation vectors, the temperature control strategy optimization system needs to evaluate the first number of target representation vectors according to the trigger prediction reference value. Specifically, a trigger prediction probability is calculated for each target representation vector, which represents the likelihood of the candidate temperature control strategy corresponding to the target representation vector being triggered.

[0019] To calculate the trigger prediction probability, a feasible implementation is to use a machine learning model, such as a logistic regression model, to predict the probability of an event occurring based on the input target representation vector. The formula is: where P(Y=1|X) represents the probability of the target representation vector X being triggered. the probability that the candidate temperature control strategy is triggered (Y = 1), is a parameter of the model, which needs to be estimated by training data. Another way is to use a decision tree model. The decision tree model makes decisions by building a tree structure, where each internal node represents a test on a feature, each branch represents a test output, and each leaf node represents a class or value. The temperature control strategy optimization system can determine the trigger prediction probability of the target representation vector according to its path in the decision tree. After calculating the trigger prediction probability of each target representation vector, compare the probabilities with the trigger prediction reference value. If the trigger prediction probability of a target representation vector is not less than the trigger prediction reference value, the target representation vector will be selected as the target mixed representation vector; otherwise, if the trigger prediction probability is less than the trigger prediction reference value, the target representation vector will be filtered out.

[0020] Step S103: loading the second number of target mixed representation vectors into the mixed trigger prediction network respectively, and outputting the second number of pending mixed trigger prediction probabilities based on the mixed trigger prediction network.

[0021] The mixed trigger prediction network is a trained neural network model, which is used to predict the probability of the corresponding candidate electric vehicle battery temperature control strategy being triggered according to the input target mixed representation vector. The structure and parameters of the network are determined in the training process, and a large amount of sample data is learned to make the network accurately capture the relationship between the target mixed representation vector and the strategy trigger probability.

[0022] The second number of target mixed representation vectors are selected from the first number of target representation vectors according to the trigger prediction reference value in step S102, and the corresponding candidate temperature control strategies have a relatively high probability of being triggered. For example, after screening, 5 target mixed representation vectors are obtained, which correspond to 5 different candidate electric vehicle battery temperature control strategies.

[0023] The temperature control strategy optimization system inputs the second number of target mixed representation vectors into the mixed trigger prediction network one by one. During the input process, each target mixed representation vector is transmitted and calculated between the neurons of each layer of the network. The mixed trigger prediction network is composed of an input layer, a hidden layer and an output layer, for example. The input layer receives the target mixed representation vector, the hidden layer performs nonlinear transformation and feature extraction on the input information, and the output layer outputs the pending mixed trigger prediction probability corresponding to the target mixed representation vector.

[0024] In actual calculation, each layer of neurons of the mixed trigger prediction network has corresponding weights and biases. Taking a fully connected neural network as an example, assuming that the hidden layer has n neurons, and the target mixed representation vector of the input layer is , the input of the i-th hidden layer neuron is wherein is the weight from the j-th neuron in the input layer to the i-th neuron in the hidden layer, is the bias of the i-th neuron in the hidden layer. Then, the output of the i-th neuron in the hidden layer is obtained by processing through an activation function A feasible activation function is the Sigmoid function , the ReLU function f(x) = max(0, x), etc. Next, the output of the hidden layer is taken as the input of the output layer, and the same weighted summation and activation function processing are performed, and finally the to-be-determined hybrid trigger prediction probability is obtained.

[0025] Step S104: determining a third number of target hybrid trigger prediction probabilities from the second number of to-be-determined hybrid trigger prediction probabilities according to a hybrid trigger prediction reference value, wherein the third number is a positive integer not greater than the second number.

[0026] The hybrid trigger prediction reference value is a pre-set threshold value, which is an indicator for further filtering strategies that are less likely to be triggered. Similar to the trigger prediction reference value, the setting of the hybrid trigger prediction reference value needs to consider various factors, such as the performance requirements of the electric vehicle, the characteristics of the battery, historical data, etc. For example, if the hybrid trigger prediction reference value is set to 0.6, only the strategies with a to-be-determined hybrid trigger prediction probability greater than or equal to 0.6 are likely to be retained. The second number of to-be-determined hybrid trigger prediction probabilities is the output result obtained after the second number of target hybrid representation vectors are loaded into the hybrid trigger prediction network in step S103. Each to-be-determined hybrid trigger prediction probability represents the likelihood of the corresponding candidate electric vehicle battery temperature control strategy being triggered. Assuming that 8 to-be-determined hybrid trigger prediction probabilities are obtained in step S103, corresponding to 8 different candidate temperature control strategies. The temperature control strategy optimization system compares the second number of to-be-determined hybrid trigger prediction probabilities one by one with the hybrid trigger prediction reference value. If a to-be-determined hybrid trigger prediction probability is greater than or equal to the hybrid trigger prediction reference value, the candidate temperature control strategy corresponding to the probability is selected, and the probability corresponding to the candidate temperature control strategy becomes the target hybrid trigger prediction probability; otherwise, if the to-be-determined hybrid trigger prediction probability is less than the hybrid trigger prediction reference value, the candidate temperature control strategy is excluded.

[0027] The temperature control strategy optimization system can be implemented by using sorting and comparison techniques. First, the temperature control strategy optimization system can sort the second number of pending mixed trigger prediction probabilities, for example, in descending order. In this way, in the comparison process, the temperature control strategy optimization system can start with the probability with the largest value, and once a probability less than the mixed trigger prediction reference value is encountered, the subsequent comparison can be stopped because the probabilities that follow will only be smaller. Sorting can use feasible sorting algorithms, such as quicksort, mergesort, etc. The average time complexity of quicksort is O(n log n), and by selecting a benchmark element, the array is divided into two parts, so that the elements on the left part are less than or equal to the benchmark element, and the elements on the right part are greater than or equal to the benchmark element, and then the left and right parts are sorted recursively.

[0028] Suppose the eight pending mixed trigger prediction probabilities obtained through step S103 are 0.2, 0.4, 0.6, 0.7, 0.8, 0.3, 0.5, and 0.9, and the mixed trigger prediction reference value is 0.6. The temperature control strategy optimization system first sorts these probabilities to obtain 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, and 0.2. Then, starting with the probability with the largest value, 0.9, it is found that 0.9 is greater than 0.6, 0.8 is greater than 0.6, 0.7 is greater than 0.6, 0.6 is equal to 0.6, and 0.5 is less than 0.6, at which point the temperature control strategy optimization system stops comparing. Then, the four probabilities 0.9, 0.8, 0.7, and 0.6 corresponding to the candidate temperature control strategies are selected, and their corresponding probabilities become the target mixed trigger prediction probabilities, and the third number is 4. Through the screening of step S104, the temperature control strategy optimization system can further narrow the range of candidate temperature control strategies and focus on those strategies that are more likely to be triggered. This not only reduces the complexity of subsequent calculations, but also improves the accuracy and effectiveness of the final determined temperature control strategy for electric vehicles.

[0029] Step S105: According to the third number of target mixed trigger prediction probabilities corresponding to the candidate electric vehicle battery temperature control strategies, respectively, the third number of target mixed trigger prediction probabilities are weighted, and the candidate electric vehicle battery temperature control strategy corresponding to the largest target mixed trigger prediction probability after weighting is selected as the electric vehicle temperature control strategy.

[0030] The third number of target mixed trigger prediction probabilities are selected from the second number of pending mixed trigger prediction probabilities according to the mixed trigger prediction reference value in step S104, and the candidate temperature control strategies corresponding to these probabilities have a high probability of being triggered. Each target mixed trigger prediction probability represents the probability of the corresponding candidate temperature control strategy being triggered under the current conditions.

[0031] The predetermined weighting coefficients of the candidate electric vehicle battery temperature control strategies are weight values set for each strategy after considering various factors. These factors include but are not limited to energy consumption, impact on battery life, cost, etc. For example, for an active cooling strategy, if its energy consumption is high but it can quickly and effectively reduce the battery temperature, a relatively low weighting coefficient can be set for it; while for a passive cooling strategy, although the cooling speed may be slow, the energy consumption is low and the impact on the battery life is small, a relatively high weighting coefficient can be set for it. Assuming that 3 target hybrid trigger prediction probabilities are obtained after step S104 screening, corresponding to 3 candidate temperature control strategies, their weighting coefficients are 0.2, 0.3, and 0.5 respectively.

[0032] The process of weighting the third number of target hybrid trigger prediction probabilities by the temperature control strategy optimization system is to multiply each target hybrid trigger prediction probability by its corresponding weighting coefficient. Assuming that the 3 target hybrid trigger prediction probabilities are 0.7, 0.8, and 0.6 respectively, the weighted results are 0.7x0.2=0.14, 0.8x0.3=0.24, and 0.6x0.5=0.3 respectively. In this process, the weighting coefficient plays a role in adjusting the importance of each strategy, so that different strategies have different proportions in the final decision.

[0033] After completing the weighting calculation, the weighted results are compared to find the maximum value. The candidate temperature control strategy corresponding to this maximum value is the final determined electric vehicle temperature control strategy. In the above example, the maximum value after weighting is 0.3, and the candidate temperature control strategy corresponding to it will be used as the electric vehicle temperature control strategy by the temperature control strategy optimization system.

[0034] Through this weighted comparison method, the temperature control strategy optimization system can consider the trigger possibility of the strategy and the comprehensive performance of each strategy, so as to select the most suitable temperature control strategy for the current situation. This method not only considers the probability of the strategy being triggered, but also considers various factors in the actual application of the strategy, so that the finally determined temperature control strategy is more scientific and reasonable.

[0035] As an implementation manner, the training process of the hybrid trigger prediction network includes the following steps: step S10: performing feature representation on each strategy trigger monitoring sequence sample in the first number of strategy trigger monitoring sequence samples to obtain a sample feature vector corresponding to each strategy trigger monitoring sequence sample, wherein each sample feature vector carries a corresponding sample strategy type label.

[0036] The first number of policy trigger monitoring sequence samples are a dataset for training the hybrid trigger prediction network, which record the triggering conditions of different candidate electric vehicle battery temperature control policies. Each policy trigger monitoring sequence sample contains a series of information related to policy triggering, such as trigger time, battery temperature at trigger time, environment temperature, battery power, etc.

[0037] Feature representation is the process of converting policy trigger monitoring sequence samples into sample feature vectors. The temperature control policy optimization system needs to extract key features from policy trigger monitoring sequence samples and combine these features into a vector. When performing feature representation, the temperature control policy optimization system can use various technical means.

[0038] For categorical features, such as the execution type of the policy (active cooling, passive cooling), the temperature control policy optimization system can use One-Hot Encoding for processing. One-Hot Encoding converts each category into a binary vector, and the length of the vector is equal to the number of categories. Only the position corresponding to the category is 1, and the rest is 0. Assuming there are two execution types, active cooling is represented by the vector [1, 0], and passive cooling is represented by the vector [0, 1].

[0039] For numerical features, such as battery temperature, environment temperature, and battery power, since the value range of these features may differ greatly, in order to avoid some features having too much influence on the model, normalization processing is needed. The feasible normalization methods include Min-Max Normalization and Z-score normalization.

[0040] After completing feature extraction and processing, the temperature control policy optimization system combines these features into a vector, which is the sample feature vector. Each dimension of the sample feature vector corresponds to a feature, and its length depends on the number of selected features. Assuming that battery temperature, environment temperature, battery power, and policy execution type are selected as features, after processing, the sample feature vector may be a four-dimensional vector, such as [0.5, 0.3, 0.6, [1, 0]], where the first three dimensions correspond to the normalized battery temperature, environment temperature, and battery power, respectively, and the last dimension is the policy execution type represented by One-Hot Encoding.

[0041] Each sample feature vector also needs to carry a corresponding sample policy type label, which indicates the policy type corresponding to the sample feature vector, such as active cooling policy or passive cooling policy. The sample policy type label can be a label, such as "active" or "passive", or a numerical encoding, such as 1 for active policy and 0 for passive policy.

[0042] Step S20: determining, according to the trigger prediction reference value, a second number of mixed sample feature vectors from the first number of sample feature vectors, wherein the second number is a positive integer not greater than the first number.

[0043] In step S20, the temperature control strategy optimization system determines, according to the trigger prediction reference value, a second number of mixed sample feature vectors from the first number of sample feature vectors, wherein the second number is a positive integer not greater than the first number. This step is a key screening link in the training process of the mixed trigger prediction network, and its purpose is to select those vectors that are more likely to correspond to the trigger strategy from a large number of sample feature vectors, so as to reduce the calculation amount of subsequent network training, and improve the training efficiency and the accuracy of the model. The trigger prediction reference value is a pre-set threshold value for initially filtering sample feature vectors corresponding to strategies that are less likely to be triggered. The setting of the threshold value needs to consider various factors, such as the performance requirements of the electric vehicle, the characteristics of the battery, and historical data, etc. For example, if the trigger prediction reference value is set to 0.2, it means that only sample feature vectors with a trigger probability greater than or equal to 0.2 will be retained.

[0044] The first number of sample feature vectors is the result obtained after the first number of strategy trigger monitoring sequence samples are represented by features in step S10. Each sample feature vector represents the feature information of a strategy trigger monitoring sequence sample and carries a corresponding sample strategy type label. Assuming that the first number is 100, there are 100 sample feature vectors, each corresponding to a different strategy trigger monitoring sequence sample.

[0045] When determining the second number of mixed sample feature vectors, the temperature control strategy optimization system first calculates a trigger prediction probability for each sample feature vector. Specifically, the temperature control strategy optimization system can determine a strategy trigger prediction network according to the sample strategy type label, and load the sample feature vector corresponding to the sample strategy type label into the network, and output the trigger prediction probability corresponding to the sample feature vector from the network. The strategy trigger prediction network includes an active strategy trigger prediction network and a passive strategy trigger prediction network, and the trigger prediction probability includes an active trigger prediction probability and a passive trigger prediction probability.

[0046] If the sample strategy type label is an active label, an active strategy trigger prediction network is determined, and the sample feature vector corresponding to the active label is loaded into the network, and the active trigger prediction probability corresponding to the sample feature vector is output based on the active strategy trigger prediction network. For example, a sample feature vector corresponds to an active label, and the temperature control strategy optimization system inputs the vector into the active strategy trigger prediction network, and obtains an active trigger prediction probability of 0.3.

[0047] If the sample strategy type label is a passive label, a passive strategy trigger prediction network is determined, and the sample feature vector corresponding to the passive label is loaded into the network. The passive trigger prediction probability corresponding to the sample feature vector is output based on the passive strategy trigger prediction network. For example, another sample feature vector corresponds to a passive label, and the passive trigger prediction probability is 0.1 after being input into the passive strategy trigger prediction network.

[0048] After obtaining the trigger prediction probabilities of all sample feature vectors, among the first number of sample feature vectors, sample feature vectors with trigger prediction probabilities not less than the trigger prediction reference value are selected, thereby obtaining the second number of mixed sample feature vectors. For example, among the 100 sample feature vectors described above, the calculated trigger prediction probabilities are high and low. When the trigger prediction reference value is 0.2, sample feature vectors with trigger prediction probabilities greater than or equal to 0.2 are selected. Assuming that 30 sample feature vectors are obtained after screening, the second number is 30, and the 30 sample feature vectors are mixed sample feature vectors.

[0049] The temperature control strategy optimization system can use sorting and comparison techniques. First, the trigger prediction probabilities of all sample feature vectors are sorted, and then the probabilities are compared from the largest probability. Once a probability less than the trigger prediction reference value is encountered, subsequent comparisons are stopped. Sorting can use available sorting algorithms, such as quicksort, with an average time complexity of O(n log n).

[0050] Step S30: Load the second number of mixed sample feature vectors into the mixed trigger prediction network, and output the second number of mixed trigger prediction probabilities based on the mixed trigger prediction network.

[0051] The second number of mixed sample feature vectors are selected from the first number of sample feature vectors according to the trigger prediction reference value in step S20. These vectors correspond to strategies with relatively high trigger likelihood. Each mixed sample feature vector represents the feature information of a strategy trigger monitoring sequence sample and carries the corresponding sample strategy type label. For example, 50 mixed sample feature vectors are obtained after screening, corresponding to 50 different strategy trigger conditions. The mixed trigger prediction network is a designed and constructed neural network model, and its structure and parameters are randomly initialized at the beginning of training. The main function of the network is to receive mixed sample feature vectors as input, perform neuron calculation and transmission inside the network, and output the mixed trigger prediction probability corresponding to each sample. The mixed trigger prediction network, for example, consists of an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving mixed sample feature vectors, the hidden layer performs nonlinear transformation and feature extraction on the input information, and the output layer outputs the final mixed trigger prediction probability.

[0052] The temperature control strategy optimization system inputs the second number of mixed sample representation vectors into the mixed trigger prediction network in turn. During the input process, each mixed sample representation vector is transmitted and calculated between the neurons of each layer of the network. As described above, the full connection neural network.

[0053] Step S40: load the second number of mixed sample representation vectors into the collaborative trigger prediction network respectively, and output the second number of collaborative trigger prediction probabilities based on the collaborative trigger prediction network.

[0054] The second number of mixed sample representation vectors are obtained after screening in step S20. The strategies corresponding to these vectors have a relatively high possibility of being triggered, and have been used for the calculation of the mixed trigger prediction network in step S30. Each mixed sample representation vector contains key feature information of the strategy trigger monitoring sequence sample and the corresponding sample strategy type label. For example, assume that 80 mixed sample representation vectors are obtained after screening, which correspond to different electric vehicle battery temperature control strategy trigger conditions.

[0055] The collaborative trigger prediction network is a neural network model that cooperates with the mixed trigger prediction network. Its structure may be similar to or different from that of the mixed trigger prediction network, but both aim to predict the trigger probability of the strategy. The role of the collaborative trigger prediction network is to evaluate the trigger probability of the strategy from another perspective to provide additional supervision information to help the mixed trigger prediction network learn and optimize better. This network also consists of an input layer, a hidden layer, and an output layer. The input layer receives mixed sample representation vectors, the hidden layer performs feature extraction and nonlinear transformation, and the output layer outputs collaborative trigger prediction probabilities.

[0056] The temperature control strategy optimization system inputs the second number of mixed sample representation vectors into the collaborative trigger prediction network in turn. During the input process, each mixed sample representation vector is transmitted and calculated between the neurons of each layer of the collaborative trigger prediction network. Taking a full connection neural network as an example, assume that the hidden layer of the collaborative trigger prediction network has k neurons, and the mixed sample representation vector received by the input layer is The input of the pth hidden layer neuron is where is the weight from the qth neuron of the input layer to the pth neuron of the hidden layer, is the bias of the pth neuron of the hidden layer. Then, the output d p of the pth neuron of the hidden layer is obtained by processing h p through an activation function, g(h p ). A feasible activation function is the Sigmoid function ReLU function g(x) = max(0, x), etc. Then, the output of the hidden layer is taken as the input of the output layer, and the weighted sum and activation function processing are also performed, and finally the synergistic trigger prediction probability is obtained.

[0057] The temperature control strategy optimization system performs such processing on the second number of mixed sample representation vectors, thereby obtaining the second number of synergistic trigger prediction probabilities. These synergistic trigger prediction probabilities are correlated with the mixed trigger prediction probabilities output by the mixed trigger prediction network in step S30. By comparing the two sets of probabilities, differences and deficiencies in the network's prediction of strategy trigger probabilities can be found. For example, if the mixed trigger prediction probability of a certain strategy is 0.6, while the synergistic trigger prediction probability is 0.8, it indicates that the two networks have different judgments on the possibility of the strategy triggering. Subsequently, based on this difference, combined with the uniformity supervision requirements, etc., the supervision cost can be calculated to adjust the parameters of the mixed trigger prediction network, so that the mixed trigger prediction network can more accurately predict the trigger probability of the strategy, improve the performance and reliability of the entire temperature control strategy optimization system, and achieve more scientific and reasonable electric vehicle battery temperature control.

[0058] Step S50: According to the uniformity supervision requirement, the second number of mixed trigger prediction probabilities and the second number of synergistic trigger prediction probabilities are obtained, and the first supervision cost is obtained.

[0059] In step S50, the temperature control strategy optimization system obtains the first supervision cost by performing cost acquisition on the second number of mixed trigger prediction probabilities and the second number of synergistic trigger prediction probabilities according to the uniformity supervision requirement. This step compares the outputs of the mixed trigger prediction network and the synergistic trigger prediction network to measure the differences between them, thereby providing a basis for adjusting the network parameters, ensuring that the trigger prediction probabilities of different strategies are consistent or similar in certain statistical quantities, and improving the accuracy and stability of the mixed trigger prediction network.

[0060] The uniformity supervision requirement is a constraint condition used to ensure that the trigger prediction probabilities of different strategies are consistent or similar in certain statistical quantities. This requirement helps to avoid excessive bias in the network's prediction of different types of strategies, making the network's output probabilities more consistent with actual situations. For example, the uniformity supervision requirement can be the uniformity supervision of the average probability, that is, the mixed trigger prediction network and the synergistic trigger prediction network are required to have similar average values of the trigger prediction probabilities of different types of strategies.

[0061] The second number of mixed trigger prediction probabilities are the results outputted by the mixed trigger prediction network in step S30 after processing the second number of mixed sample feature vectors, each probability representing the likelihood of the corresponding strategy being triggered. The second number of cooperative trigger prediction probabilities are the results outputted by the cooperative trigger prediction network in step S40 after processing the same second number of mixed sample feature vectors. These two sets of probabilities are the basic data for calculating the first supervision cost.

[0062] Suppose the second number is 50, i.e., there are 50 mixed trigger prediction probabilities and 50 cooperative trigger prediction probabilities. Among them, the mixed trigger prediction probabilities are , and the cooperative trigger prediction probabilities are .

[0063] If the uniformity supervision requirement is the uniformity supervision of the average of the probabilities, the temperature control strategy optimization system needs to obtain the cost according to the following steps. First, according to the sample strategy type label, the mixed trigger prediction probabilities and the cooperative trigger prediction probabilities are divided into active and passive two categories. Suppose that among the 50 samples, there are 30 active samples and 20 passive samples. The active mixed trigger prediction probability is , the active cooperative trigger prediction probability is ; the passive mixed trigger prediction probability is , and the passive cooperative trigger prediction probability is .

[0064] Then, the mean values of each category are calculated. The active mixed mean value is , the passive mixed mean value is , the active cooperative mean value is , and the passive cooperative mean value is .

[0065] Finally, the mean values are used to obtain the cost. A feasible cost function is the mean square error (MSE), whose formula is , y i is the true value, is the predicted value. In this case, the first supervision cost C can be represented as .

[0066] By calculating the first supervision cost, the difference between the outputs of the hybrid trigger prediction network and the collaborative trigger prediction network can be quantified. This cost reflects the performance of the networks in meeting the uniformity supervision requirements. If the first supervision cost is large, it means that the outputs of the two networks differ greatly, and the parameters of the hybrid trigger prediction network need to be adjusted to reduce the difference and make the network output more consistent with the uniformity supervision requirements. For example, a gradient descent algorithm can be used to update the parameters of the hybrid trigger prediction network according to the first supervision cost, gradually reducing the cost and improving the performance of the network. In practical applications, different cost functions can be selected flexibly to obtain the cost according to different uniformity supervision requirements. In addition to mean square error, other cost functions such as cross-entropy loss function can also be used. At the same time, in order to improve the calculation efficiency and accuracy, batch processing can be used to calculate multiple samples together. In this way, the temperature control strategy optimization system can more effectively train and optimize the hybrid trigger prediction network, making it better predict the trigger probability of the strategy and provide more reliable support for the optimization of the temperature control strategy of the electric vehicle.

[0067] Step S60: Adjust the network parameters of the hybrid trigger prediction network according to the first supervision cost.

[0068] In step S60, the temperature control strategy optimization system adjusts the network parameters of the hybrid trigger prediction network according to the first supervision cost. This step adjusts the network parameters so that the output of the hybrid trigger prediction network can better meet the uniformity supervision requirements, thereby improving the accuracy of the network in predicting the trigger probability of the strategy.

[0069] The first supervision cost is the result obtained by calculating the cost of the second number of hybrid trigger prediction probabilities and the second number of collaborative trigger prediction probabilities according to the uniformity supervision requirements in step S50. It reflects the difference between the outputs of the hybrid trigger prediction network and the collaborative trigger prediction network, and reflects the degree of conformity of the hybrid trigger prediction network to the uniformity supervision requirements under the current network parameters. For example, if the first supervision cost is large, it means that the outputs of the two networks differ significantly, and the hybrid trigger prediction network performs poorly in meeting the uniformity supervision requirements; on the contrary, if the first supervision cost is small, it means that the network output is more consistent with the requirements.

[0070] The network parameters of the hybrid trigger prediction network refer to the weights and biases and other parameters in the network, which determine the calculation method and output result of the network. At the beginning of training, the network parameters are randomly initialized, and as the training progresses, they need to be adjusted according to the first supervision cost to make the network gradually converge to the optimal state.

[0071] The temperature control strategy optimization system can adjust the network parameters of the hybrid trigger prediction network by using a gradient descent algorithm to update the network parameters along the negative gradient direction of the first supervision cost, so as to gradually reduce the value of the loss function.

[0072] Suppose the network parameters of the hybrid trigger prediction network are , and the first supervision cost is . The update formula of the gradient descent algorithm is: , wherein is the current network parameter, is the updated network parameter, is the learning rate, which controls the step size of each update, is the first supervision cost at the gradient .

[0073] In each iteration, the temperature control strategy optimization system calculates the gradient according to the current first supervision cost and updates the network parameters. As the number of iterations increases, the first supervision cost gradually decreases, and the network parameters gradually converge to the optimal value. When the first supervision cost is less than a certain preset threshold or reaches the maximum number of iterations, the training process ends.

[0074] By continuously adjusting the network parameters of the hybrid trigger prediction network, the network can better learn the rules of policy triggering and improve the prediction accuracy of policy trigger probability. At the same time, meeting the uniformity supervision requirement can ensure that the prediction results of the network for different types of policies are consistent in statistics, avoiding over-prediction or underestimation of certain types of policies, thereby providing more reliable basis for the optimization of electric vehicle temperature control and achieving more scientific and reasonable battery temperature control.

[0075] As an implementation, after obtaining the first supervision cost according to the uniformity supervision requirement, the method further includes: step S501: obtaining a dispersion cost by obtaining a dispersion cost according to the dispersion supervision requirement.

[0076] Based on this, step S60 adjusts the network parameters of the hybrid trigger prediction network according to the first supervision cost, including: step S61: adjusting the network parameters of the hybrid trigger prediction network according to the first supervision cost and the dispersion cost.

[0077] In step S501, the temperature control strategy optimization system obtains the dispersion cost according to the dispersion supervision requirement and the second number of mixed trigger prediction probabilities. In step S61, the network parameters of the mixed trigger prediction network are adjusted according to the first supervision cost and the dispersion cost. The distribution characteristics of the network output probability are further constrained, and the network parameters are optimized by comprehensively considering various cost information, so as to improve the performance and prediction accuracy of the network.

[0078] The dispersion supervision requirement is a difference constraint used to ensure that the different probabilities of the network output meet certain standards or expectations in statistics. It supplements the shortcomings of the uniformity supervision requirement and avoids the concentration or lack of diversity of the network output probability. For example, in the scenario of electric vehicle temperature control strategy, if all the strategy trigger probabilities of the network output are very close, it is difficult to effectively distinguish the advantages and disadvantages of different strategies, which is not conducive to selecting a suitable temperature control strategy. Through the dispersion supervision requirement, the probability distribution of the network output can be made more reasonable, and the effectiveness of strategy selection can be improved.

[0079] In step S501, the dispersion cost is obtained by obtaining the cost of the second number of mixed trigger prediction probabilities. Assuming that the second number is N, the mixed trigger prediction probabilities are According to the sample strategy type label, these probabilities can be divided into active and passive types. Let the number of active samples be N a , the active mixed trigger prediction probability be , the number of passive samples be N p , the passive mixed trigger prediction probability be , and .

[0080] To calculate the dispersion cost, the dispersion (standard deviation) and mean of each type need to be calculated first. The active mixed dispersion can be calculated by the formula , where is the active mixed mean, is the i-th active mixed trigger prediction probability. The passive mixed dispersion can be calculated by the formula , where is the passive mixed mean, is the i-th passive mixed trigger prediction probability.

[0081] After obtaining the active mixed dispersion, the active mixed mean, the passive mixed dispersion, and the passive mixed mean, the temperature control strategy optimization system needs to obtain the cost of these values to obtain the dispersion cost. One feasible way is to use the idea of mean square error to construct a cost function. Let the dispersion cost be , then , where These are the expected active mixture dispersion, active mixture mean, passive mixture dispersion, and passive mixture mean, respectively. These are the corresponding weighting coefficients, used to adjust the importance of each item in the dispersion cost. For example, if you want to focus more on the constraint of dispersion, you can set w1 and w3 to be larger.

[0082] After obtaining the dispersion cost, in step S61, the temperature control strategy optimization system adjusts the network parameters of the hybrid trigger prediction network based on the first supervision cost and the dispersion cost. The first supervision cost is calculated in step S50 according to the uniformity supervision requirement; it reflects the degree of difference between the outputs of the hybrid trigger prediction network and the cooperative trigger prediction network. Let the first supervision cost be... The cost of dispersion is Comprehensive cost It can be obtained by weighted summation, that is... ,in These are weighting coefficients used to balance the importance of the first supervision cost and the dispersion cost. For example, if uniformity supervision requirements are given more weight, then... The setting is too large.

[0083] When adjusting the network parameters of the hybrid trigger prediction network in a temperature control strategy optimization system, a gradient descent algorithm is feasible. Assume the network parameters of the hybrid trigger prediction network are... The update formula for the gradient descent algorithm is: ,in These are the current network parameters. These are the updated network parameters. It's the learning rate, which controls the step size for each update. It is a comprehensive cost exist The gradient at that point.

[0084] In actual gradient calculation, due to the overall cost It is the first supervisory cost C u and the cost of dispersion C d The weighted result, based on the linear property of differentiation, This requires calculating the gradients of the first supervision cost and the distributed cost with respect to the network parameters separately, and then combining them according to the weights.

[0085] In each iteration, the temperature control strategy optimization system first calculates the first supervised cost C under the current network parameters. u and the cost of dispersion C d Then, calculate the overall cost C according to the above formula. total Next, based on automatic differentiation calculation... and update the network parameters according to the update formula of the gradient descent algorithm. With the increase of the number of iterations, the comprehensive cost C total will gradually decrease, and the network parameters will gradually converge to the optimal value.

[0086] When the comprehensive cost C total is less than a certain preset threshold or reaches the maximum number of iterations, the training process ends. In this way, the temperature control strategy optimization system can comprehensively consider the uniformity supervision requirement and the dispersion supervision requirement, and effectively adjust the network parameters of the hybrid trigger prediction network.

[0087] Through the above comprehensive adjustment method, the hybrid trigger prediction network can not only maintain a certain consistency in the trigger prediction probabilities of different types of strategies (meet the uniformity supervision requirement), but also ensure that the output probabilities have reasonable dispersion (meet the dispersion supervision requirement). In the application scenario of electric vehicle temperature control strategy, such network can more accurately predict the trigger probability of different temperature control strategies, providing a more reliable basis for selecting the appropriate temperature control strategy. For example, when the network output of different strategy trigger probabilities has a reasonable distribution and is consistent in statistics, the system can more accurately determine which temperature control strategy should be used in different situations, thereby achieving more scientific and efficient battery temperature control, which is beneficial to the health maintenance and life extension of the battery.

[0088] As an implementation manner, if the uniformity supervision requirement is a probability average value; then in step S50, the second number of mixed trigger prediction probabilities and the second number of cooperative trigger prediction probabilities are subjected to cost acquisition according to the uniformity supervision requirement, to obtain a first supervision cost, including: in step S51, according to the sample strategy type label, in the second number of mixed trigger prediction probabilities, determining the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability; in step S52, according to the sample strategy type label, in the second number of cooperative trigger prediction probabilities, determining the active cooperative trigger prediction probability corresponding to the active sample strategy type label, the active cooperative number corresponding to the active cooperative trigger prediction probability, the passive cooperative trigger prediction probability corresponding to the passive sample strategy type label, and the passive cooperative number corresponding to the passive cooperative trigger prediction probability; in step S53, the active mixed mean value is acquired according to the active mixed trigger prediction probability and the active mixed number, and the passive mixed mean value is acquired according to the passive mixed trigger prediction probability and the passive mixed number; in step S54, the active cooperative mean value is acquired according to the active cooperative trigger prediction probability and the active cooperative number, and the passive cooperative mean value is acquired according to the passive cooperative trigger prediction probability and the passive cooperative number; in step S55, the active mixed mean value, the passive mixed mean value, the active cooperative mean value, and the passive cooperative mean value are subjected to cost acquisition, to obtain the first supervision cost.

[0089] Steps S51 to S55 compare the average values of the trigger prediction probabilities of different types of strategies (active and passive) in the mixed trigger prediction network and the cooperative trigger prediction network, quantify the differences between the outputs of the two networks, provide a basis for subsequent adjustment of network parameters of the mixed trigger prediction network, and thus ensure that the network can more accurately predict the trigger probabilities of different types of strategies, to meet the needs of electric vehicle temperature control strategy optimization.

[0090] In step S51, according to the sample strategy type label, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability are determined from the second number of mixed trigger prediction probabilities. The sample strategy type label is marked for each sample feature vector in step S10, which is used to distinguish whether the strategy is an active type or a passive type, such as an active cooling strategy or a passive cooling strategy. The second number of mixed trigger prediction probabilities is the result output by the mixed trigger prediction network after processing the second number of mixed sample feature vectors in step S30. The temperature control strategy optimization system divides each mixed trigger prediction probability into active and passive categories by checking the sample strategy type label corresponding to each mixed trigger prediction probability.

[0091] In step S52, according to the sample strategy type label, the active cooperative trigger prediction probability corresponding to the active sample strategy type label, the active cooperative number corresponding to the active cooperative trigger prediction probability, the passive cooperative trigger prediction probability corresponding to the passive sample strategy type label, and the passive cooperative number corresponding to the passive cooperative trigger prediction probability are determined from the second number of cooperative trigger prediction probabilities. The second number of cooperative trigger prediction probabilities is the result output by the cooperative trigger prediction network after processing the same second number of mixed sample feature vectors in step S40. Similarly, the temperature control strategy optimization system classifies the cooperative trigger prediction probabilities according to the sample strategy type label. Following the above example, among the 100 cooperative trigger prediction probabilities, 60 correspond to the active sample strategy type label, and these probabilities are the active cooperative trigger prediction probabilities, with an active cooperative number of 60; 40 correspond to the passive sample strategy type label, and these probabilities are the passive cooperative trigger prediction probabilities, with a passive cooperative number of 40.

[0092] In step S53, the active mixed mean is obtained according to the active mixed trigger prediction probability and the active mixed number, and the passive mixed mean is obtained according to the passive mixed trigger prediction probability and the passive mixed number. The mean is the average value of a set of data, which can reflect the central tendency of the set of data. The calculation formula of the active mixed mean is wherein is the active mixed number, is the i-th active mixed trigger prediction probability. The calculation formula of the passive mixed mean is wherein is the passive mixed number, is the i-th passive mixed trigger prediction probability.

[0093] In step S54, the active cooperation mean value is obtained according to the active cooperation trigger prediction probability and the active cooperation number, and the passive cooperation mean value is obtained according to the passive cooperation trigger prediction probability and the passive cooperation number. The calculation formula of the active cooperation mean value is where m a is the active cooperation number, Q ai is the ith active cooperation trigger prediction probability. The calculation formula of the passive cooperation mean value is where m p is the passive cooperation number, is the ith passive cooperation trigger prediction probability.

[0094] In step S55, the temperature control strategy optimization system obtains the first supervision cost by performing cost acquisition on the active mixture mean value, the passive mixture mean value, the active cooperation mean value, and the passive cooperation mean value. The first supervision cost is an index for measuring the difference between the mixed trigger prediction network and the cooperation trigger prediction network in the average value of the trigger prediction probability of different types of strategies. A feasible cost acquisition method is to use the mean square error (MSE), and the formula is where y i is the true value, is the predicted value. In this case, the first supervision cost C can be expressed as This formula comprehensively considers the mean value difference between the active and passive types of strategies in the mixed trigger prediction network and the cooperation trigger prediction network.

[0095] As an embodiment, if the dispersion supervision requirement is the dispersion supervision of the probability distribution, then in step S501, the dispersion cost is obtained by performing cost acquisition on the second number of mixed trigger prediction probabilities according to the dispersion supervision requirement, including: in step S5011, the active mixture dispersion is obtained according to the active mixture trigger prediction probability and the active mixture number, and the passive mixture dispersion is obtained according to the passive mixture trigger prediction probability and the passive mixture number; in step S5012, the active cooperation dispersion is obtained according to the active cooperation trigger prediction probability and the active cooperation number, and the passive cooperation dispersion is obtained according to the passive cooperation trigger prediction probability and the passive cooperation number; in step S5013, the dispersion cost is obtained by performing cost acquisition on the active mixture dispersion, the active mixture mean value, the passive mixture dispersion, the passive mixture mean value, the active cooperation dispersion, the active cooperation mean value, the passive cooperation dispersion, and the passive cooperation mean value.

[0096] In step S5011, the active mixing dispersion is obtained according to the active mixing trigger prediction probability and the active mixing number, and the passive mixing dispersion is obtained according to the passive mixing trigger prediction probability and the passive mixing number. The dispersion, also known as the standard deviation, is a statistical quantity used to measure the dispersion degree of a group of data or the dispersion of the distribution, which reflects the deviation of the data from the mean. The active mixing trigger prediction probability is the probability corresponding to the active sample strategy type label selected from the second number of mixing trigger prediction probabilities in step S51, and the active mixing number is the number of active mixing trigger prediction probabilities. The passive mixing trigger prediction probability and the passive mixing number are the same. The calculation formula of the active mixing dispersion and the passive mixing dispersion can be referred to the foregoing content, which will not be described here.

[0097] In step S5012, the active synergy dispersion is obtained according to the active synergy trigger prediction probability and the active synergy number, and the passive synergy dispersion is obtained according to the passive synergy trigger prediction probability and the passive synergy number. The active synergy trigger prediction probability is the probability corresponding to the active sample strategy type label selected from the second number of synergy trigger prediction probabilities in step S52, and the active synergy number is the number of active synergy trigger prediction probabilities. The passive synergy trigger prediction probability and the passive synergy number are the same. The calculation formula of the active synergy dispersion is , where m a is the active synergy number, Q ai is the ith active synergy trigger prediction probability, is the active synergy mean. The calculation formula of the passive synergy dispersion is , where m p is the passive synergy number, Q pi is the ith passive synergy trigger prediction probability, is the passive synergy mean.

[0098] In step S5013, the cost of the active mixing dispersion, the active mixing mean, the passive mixing dispersion, the passive mixing mean, the active synergy dispersion, the active synergy mean, the passive synergy dispersion, and the passive synergy mean is obtained, and the dispersion cost is obtained. The dispersion cost is an index for measuring whether the dispersion degree of the probability distribution output by the mixing trigger prediction network on different types of strategies meets the requirements.

[0099] In order to calculate the dispersion cost, the temperature control strategy optimization system can construct a cost function, for example, in the form of weighted mean square error. Let the dispersion cost be , which can be expressed as , where respectively are the expected active mixing dispersion, the expected active mixing mean, the expected passive mixing dispersion, the expected passive mixing mean, the expected active synergy dispersion, the expected active synergy mean, the expected passive synergy dispersion and the expected passive synergy mean, are the corresponding weight coefficients for adjusting the importance of each term in the dispersion cost.

[0100] As an implementation, in step S20, the second number of mixed sample feature vectors are determined from the first number of sample feature vectors according to the trigger prediction reference value, including: step S21, determining a policy trigger prediction network according to the sample policy type label, and loading the sample feature vector corresponding to the sample policy type label into the policy trigger prediction network, and outputting the trigger prediction probability corresponding to the sample feature vector based on the policy trigger prediction network; step S22, selecting the sample feature vector whose trigger prediction probability is not less than the trigger prediction reference value from the first number of sample feature vectors to obtain the second number of mixed sample feature vectors.

[0101] Steps S21-S22 screen the sample feature vectors, preliminarily filter out the sample feature vectors corresponding to the policies that are less likely to be triggered, thereby reducing the calculation amount of subsequent mixed trigger prediction network training, improving the training efficiency, and at the same time making the training more focused on the policies that have greater triggering possibility, and improving the accuracy of the model.

[0102] In step S21, the temperature control policy optimization system determines a policy trigger prediction network according to the sample policy type label, and loads the sample feature vector corresponding to the sample policy type label into the policy trigger prediction network, and outputs the trigger prediction probability corresponding to the sample feature vector based on the network. The sample policy type label is the information labeled for each sample feature vector in step S10, which is used to distinguish whether the policy is an active type or a passive type, such as an active cooling policy or a passive cooling policy. The policy trigger prediction network includes an active policy trigger prediction network and a passive policy trigger prediction network, and the two networks are respectively used to predict the trigger probability of the active policy and the passive policy.

[0103] Specifically, if the sample policy type label is an active label, an active policy trigger prediction network is determined, and the sample feature vector corresponding to the active label is loaded into the network, and the active trigger prediction probability corresponding to the sample feature vector is output based on the active policy trigger prediction network. For example, there are 100 sample feature vectors, and the sample policy type label of 50 sample feature vectors is an active label. For one of the sample feature vectors X a1, which is input into the active policy trigger prediction network. The active policy trigger prediction network is a pre-trained neural network model, which can include an input layer, a hidden layer and an output layer. Assume that the input layer has n neurons corresponding to n feature dimensions of the sample feature vector, and the hidden layer has m neurons. The sample feature vector After being input into the input layer, the input of the jth neuron of the hidden layer is , where is the weight from the ith neuron of the input layer to the jth neuron of the hidden layer, is the bias of the jth neuron of the hidden layer, is an element in , and is the ith element in . After being processed by an activation function (such as a Sigmoid function ), the output of the jth neuron of the hidden layer is . The output layer receives the output of the hidden layer, and after being processed by weighted summation and an activation function, outputs the active trigger prediction probability .

[0104] If the sample policy type label is a passive label, the passive policy trigger prediction network is determined, and the sample feature vector corresponding to the passive label is loaded into the network, and the passive trigger prediction probability corresponding to the sample feature vector is output based on the passive policy trigger prediction network. For example, the sample policy type labels of the other 50 sample feature vectors are passive labels, and for the sample feature vector X p1 , the temperature control policy optimization system inputs it into the passive policy trigger prediction network, and after a similar network calculation process, outputs the passive trigger prediction probability P p1 .

[0105] In step S22, the temperature control policy optimization system selects, from the first number of sample feature vectors, sample feature vectors with trigger prediction probabilities not less than a trigger prediction reference value, to obtain a second number of mixed sample feature vectors. The trigger prediction reference value is a pre-set threshold value for preliminarily judging whether a policy corresponding to a sample feature vector has enough possibility to be triggered. For example, if the trigger prediction reference value is set to 0.3, only when the trigger prediction probability (active trigger prediction probability or passive trigger prediction probability) corresponding to the sample feature vector is greater than or equal to 0.3, the sample feature vector will be selected to become a mixed sample feature vector.

[0106] Continuing with the above example, among the 100 sample feature vectors, the 50 active sample feature vectors obtain active trigger prediction probabilities through the active policy trigger prediction network, and the 50 passive sample feature vectors obtain passive trigger prediction probabilities Compare these trigger prediction probabilities with the trigger prediction reference values ​​one by one. Assume... =0.2, which is less than the trigger prediction reference value of 0.3, then the sample representation vector It will not be selected; if =0.4, which is greater than the trigger prediction reference value of 0.3, then the sample representation vector They will be selected. The same filtering method applies to the passively triggered prediction probability. After filtering, assuming that 30 sample representation vectors are finally selected, the second quantity is 30, and these 30 sample representation vectors are the mixed sample representation vectors.

[0107] As one implementation, the policy trigger prediction network includes an active policy trigger prediction network and a passive policy trigger prediction network, and the trigger prediction probability includes an active trigger prediction probability and a passive trigger prediction probability. Based on this, step S21, determining the policy trigger prediction network according to the sample policy type label, loading the sample representation vector corresponding to the sample policy type label into the policy trigger prediction network, and outputting the trigger prediction probability corresponding to the sample representation vector based on the policy trigger prediction network, includes: Step S211: If the sample policy type label is an active label, then determining the active policy trigger prediction network, loading the sample representation vector corresponding to the active label into the active policy trigger prediction network, and outputting the active trigger prediction probability corresponding to the sample representation vector based on the active policy trigger prediction network; Step S212: If the sample policy type label is a passive label, then determining the passive policy trigger prediction network, loading the sample representation vector corresponding to the passive label into the passive policy trigger prediction network, and outputting the passive trigger prediction probability corresponding to the sample representation vector based on the passive policy trigger prediction network.

[0108] Based on this, step S22, from the first number of sample representation vectors, select sample representation vectors whose trigger prediction probability is not less than the trigger prediction reference value, to obtain the second number of mixed sample representation vectors, including: step S221: from the first number of sample representation vectors, select sample representation vectors whose active trigger prediction probability and passive trigger prediction probability are not less than the trigger prediction reference value, to obtain the second number of mixed sample representation vectors.

[0109] In step S211, if the sample strategy type label is an active label, an active strategy trigger prediction network is determined, and the sample feature vector corresponding to the active label is loaded into the network, so as to output the active trigger prediction probability corresponding to the sample feature vector. The active strategy trigger prediction network is a neural network model specially designed for active strategies, and its purpose is to predict the possibility of the active strategy being triggered according to the input sample feature vector. The sample strategy type label is the information labeled for each sample feature vector in step S10, which is used to distinguish whether the strategy is an active type or a passive type. The active label indicates that the strategy corresponding to the sample feature vector is an active strategy, such as an active cooling strategy, which requires additional energy consumption to achieve the temperature control target.

[0110] Suppose the temperature control strategy optimization system has 100 sample feature vectors, and the sample strategy type labels of 50 sample feature vectors are active labels. For one of the sample feature vectors , the temperature control strategy optimization system inputs it into the active strategy trigger prediction network. The active strategy trigger prediction network is composed of an input layer, a hidden layer, and an output layer, for example. The number of neurons in the input layer is the same as the feature dimension of the sample feature vector, which is assumed to be n. The hidden layer has m neurons for nonlinear transformation and feature extraction of the input information. The output layer has one neuron that outputs the active trigger prediction probability.

[0111] During network calculation, each neuron in the input layer receives a feature value of the sample feature vector. The input of the jth neuron in the hidden layer is which can be calculated by the formula , where is the weight from the ith neuron in the input layer to the jth neuron in the hidden layer, is the bias of the jth neuron in the hidden layer, is the ith element in the sample feature vector . In order to introduce a nonlinear factor, the is processed using an activation function, such as the Sigmoid function , and the output of the jth neuron in the hidden layer is obtained after the activation function processing . The output layer receives the output of the hidden layer, and after weighted summation and activation function processing, the active trigger prediction probability is output. Assuming that the weight of the output layer is and the bias is , the input of the output layer is , and the final active trigger prediction probability is .

[0112] In step S212, if the sample policy type is labeled as passive, a passive policy trigger prediction network is determined, and the sample representation vector corresponding to the passive label is loaded into the network, thereby outputting the passive trigger prediction probability corresponding to the sample representation vector. The passive policy trigger prediction network is a neural network model specifically designed for passive policies, used to predict the probability of a passive policy being triggered. The passive label indicates that the policy corresponding to the sample representation vector is a passive policy, such as a passive heat dissipation policy, which relies on natural environmental conditions to achieve temperature control without consuming a large amount of additional energy.

[0113] Similarly, among these 100 sample representation vectors, the sample strategy type of another 50 sample representation vectors is labeled as passive. For sample representation vectors... The temperature control strategy optimization system inputs the data into a passive policy trigger prediction network. The structure and computation process of the passive policy trigger prediction network are similar to those of the active policy trigger prediction network. The input layer receives the feature values ​​of the sample representation vector, the hidden layer processes the input information, and the output layer outputs the passive trigger prediction probability P. p1 The input to the k-th neuron in the hidden layer. , These are the weights from the i-th neuron in the input layer to the k-th neuron in the hidden layer. It is the bias of the i-th neuron in the hidden layer. It is the sample representation vector The i-th element in the hidden layer. After processing by the activation function, the output of the k-th neuron in the hidden layer is obtained. The input of the output layer The final passive trigger prediction probability .

[0114] After obtaining the active trigger prediction probability and passive trigger prediction probability of all sample representation vectors in steps S211 and S212, in step S221, the temperature control strategy optimization system selects sample representation vectors from the first number of sample representation vectors whose active trigger prediction probability and passive trigger prediction probability are not less than the trigger prediction reference value, to obtain the second number of mixed sample representation vectors. The trigger prediction reference value is a pre-set threshold used to initially filter sample representation vectors corresponding to strategies that are unlikely to be triggered.

[0115] As an implementation, if the sample strategy type label is a passive label, step S22 determines a passive strategy trigger prediction network, and loads the sample feature vector corresponding to the passive label into the passive strategy trigger prediction network. After outputting the passive trigger prediction probability corresponding to the sample feature vector based on the passive strategy trigger prediction network, the method further comprises: step S23: obtaining an active cost according to the active label and the active trigger prediction probability, and adjusting the network parameters of the active strategy trigger prediction network according to the active cost; and step S24: obtaining a passive cost according to the passive label and the passive trigger prediction probability, and adjusting the network parameters of the passive strategy trigger prediction network according to the passive cost.

[0116] Steps S23 and S24 are further operations of the temperature control strategy optimization system after obtaining the active trigger prediction probability and the passive trigger prediction probability corresponding to the sample feature vector. The core purpose is to optimize the performance of the active strategy trigger prediction network and the passive strategy trigger prediction network by calculating the cost and adjusting the network parameters, so that the two networks can more accurately predict the trigger probability of the corresponding type of strategy.

[0117] In step S23, the temperature control strategy optimization system obtains an active cost according to the active label and the active trigger prediction probability, and adjusts the network parameters of the active strategy trigger prediction network according to the active cost. The active label is a label carried by the sample feature vector to identify the active strategy corresponding to the sample feature vector. The active trigger prediction probability is the result output by the active strategy trigger prediction network after processing the sample feature vector corresponding to the active label, which represents the likelihood of the active strategy corresponding to the sample feature vector being triggered.

[0118] The active cost is an index for measuring the difference between the prediction result of the active strategy trigger prediction network and the actual situation. To obtain the active cost, a feasible cost function is, for example, mean square error (MSE), cross-entropy loss function, etc. Taking the mean square error as an example, suppose there are N sample feature vectors with active labels, and their corresponding real labels (which can be understood as the actual trigger label, if triggered, it is 1, and if not triggered, it is 0) are , and the active trigger prediction probability output by the active strategy trigger prediction network is , then the active cost C a can be calculated by the formula , where represents the real label corresponding to the sample feature vector with the active label, is the active trigger prediction probability output by the active strategy trigger prediction network after processing the sample feature vector with the active label. This formula calculates the average of the square of the difference between the true label and the prediction probability of each sample. The larger the difference, the higher the active cost, indicating that the deviation between the network's prediction result and the actual situation is larger.

[0119] After obtaining the active cost, the network parameters of the active strategy trigger prediction network are adjusted according to the cost. The network parameters include weights and biases in the network, which determine the calculation method and output result of the network. A feasible method to adjust the network parameters is the gradient descent algorithm. The basic idea of the gradient descent algorithm is to update the network parameters along the negative gradient direction of the cost function to gradually reduce the value of the cost function.

[0120] In step S24, the temperature control strategy optimization system obtains the passive cost according to the passive label and the passive trigger prediction probability, and adjusts the network parameters of the passive strategy trigger prediction network according to the passive cost. The passive label is a label carried by the sample feature vector to identify the corresponding passive strategy, and the passive trigger prediction probability is the result output by the passive strategy trigger prediction network after processing the sample feature vector corresponding to the passive label, which represents the possibility of the corresponding passive strategy being triggered.

[0121] Similarly, to obtain the passive cost, the calculation method of the aforementioned active cost can be referred to, which is not described here. After obtaining the passive cost, the network parameters of the passive strategy trigger prediction network are adjusted according to the cost.

[0122] In actual operation, steps S23 and S24 are executed multiple times. Each iteration calculates the active cost and the passive cost, and updates the network parameters of the corresponding network according to these costs. With the increase of the number of iterations, the active cost and the passive cost gradually decrease, indicating that the deviation between the network's prediction result and the actual situation is continuously reduced, and the performance of the network is continuously improved. When the active cost and the passive cost are less than a certain preset threshold, or the maximum number of iterations is reached, the training process ends.

[0123] As an implementation, after step S30, the second number of mixed sample feature vectors are loaded into the mixed trigger prediction network, and the second number of mixed trigger prediction probabilities are output based on the mixed trigger prediction network, the method further comprises: step S301: obtaining the second supervision cost by obtaining the cost of the second number of mixed trigger prediction probabilities and the trigger prediction probability according to the uniformity supervision requirement; step S302: adjusting the network parameters of the mixed trigger prediction network according to the second supervision cost.

[0124] In step S301, the temperature control strategy optimization system obtains a second supervision cost according to the uniformity supervision requirement, based on the second number of mixed trigger prediction probabilities and the trigger prediction probability. The trigger prediction probability is the result output by the strategy trigger prediction network (including the active strategy trigger prediction network and the passive strategy trigger prediction network) in step S21, representing the possibility of the strategy being triggered corresponding to the sample feature vector. The mixed trigger prediction probability is the result output by the mixed trigger prediction network after processing the second number of mixed sample feature vectors in step S30.

[0125] To obtain the second supervision cost, feasible cost functions include mean square error (MSE) and cross-entropy loss function. Assuming that the uniformity supervision requirement is the uniformity supervision of the probability average value, and the second number is N, the mixed trigger prediction probability is , and the trigger prediction probability is Taking the mean square error as an example, the second supervision cost C2 can be calculated by the formula , as described above, representing the mixed trigger prediction probability, indicating the trigger prediction probability, which calculates the average value of the square of the difference between the mixed trigger prediction probability and the trigger prediction probability of each sample. The greater the difference, the higher the second supervision cost, indicating that the difference between the output of the mixed trigger prediction network and the output of the strategy trigger prediction network in the probability average value is greater, and the degree of not meeting the uniformity supervision requirement is higher.

[0126] In actual operation, the form of the cost function can be flexibly adjusted according to different uniformity supervision requirements. If the uniformity supervision requirement is the uniformity supervision of the probability average value of a specific type of strategy (such as active strategy or passive strategy), the temperature control strategy optimization system first classifies the mixed trigger prediction probability and the trigger prediction probability according to the sample strategy type label, then calculates the cost of different types of strategies respectively, and finally obtains the second supervision cost.

[0127] In step S302, the temperature control strategy optimization system adjusts the network parameters of the mixed trigger prediction network according to the second supervision cost. The network parameters include the weights and biases in the network, which determine the calculation method and output result of the network. The purpose of adjusting the network parameters is to reduce the second supervision cost, so that the output of the mixed trigger prediction network meets the uniformity supervision requirement. A feasible adjustment method is the gradient descent algorithm. The basic idea of the gradient descent algorithm is to update the network parameters along the negative gradient direction of the cost function, so as to gradually reduce the value of the cost function.

[0128] The steps S301 and S302 are performed multiple times. Each iteration calculates the second supervision cost and updates the network parameters of the hybrid trigger prediction network according to the cost. As the number of iterations increases, the second supervision cost gradually decreases, indicating that the difference between the output of the hybrid trigger prediction network and the output of the policy trigger prediction network in meeting the uniformity supervision requirement is gradually reduced, and the performance of the network is continuously improved. When the second supervision cost is less than a certain preset threshold or reaches the maximum number of iterations, the training process ends.

[0129] Through steps S301 and S302, it can be ensured that the policy trigger probability output by the hybrid trigger prediction network and the output of the policy trigger prediction network are statistically consistent, avoiding overestimation or underestimation of some types of strategies. For example, in a high-temperature environment, the trigger probabilities of the active cooling strategy and the passive cooling strategy should maintain a reasonable relationship to some extent. If the active cooling strategy trigger probability output by the hybrid trigger prediction network is too high, and the passive cooling strategy trigger probability is too low, it may cause the system to rely too much on the active cooling strategy, increasing energy consumption. By adjusting the network parameters, the output of the hybrid trigger prediction network meets the uniformity supervision requirement, which can more accurately select the appropriate temperature control strategy, achieve more scientific and efficient battery temperature control, and be beneficial to the health maintenance and life extension of the battery.

[0130] As an implementation, in step S301, the cost of the second number of mixed trigger prediction probabilities and the trigger prediction probability is obtained according to the uniformity supervision requirement, and the second supervision cost is obtained, including: step S3011: in the second number of mixed trigger prediction probabilities, the active mixed trigger prediction probability corresponding to the active sample strategy type mark, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type mark, and the passive mixed number corresponding to the passive mixed trigger prediction probability are determined according to the sample strategy type mark; step S3012: in the first number of trigger prediction probabilities, the active trigger prediction probability corresponding to the active sample strategy type mark, the active number corresponding to the active trigger prediction probability, the passive trigger prediction probability corresponding to the passive sample strategy type mark, and the passive number corresponding to the passive trigger prediction probability are determined according to the sample strategy type mark; step S3013: the active mixed mean is obtained according to the active mixed trigger prediction probability and the active mixed number, and the passive mixed mean is obtained according to the passive mixed trigger prediction probability and the passive mixed number; step S3014: the active mean is obtained according to the active trigger prediction probability and the active number, and the passive mean is obtained according to the passive trigger prediction probability and the passive number; step S3015: if the uniformity supervision requirement is the uniformity supervision of the active probability average value, the cost of the active mixed mean, the passive mixed mean and the active mean is obtained, and the second supervision cost is obtained; step S3016: if the uniformity supervision requirement is the uniformity supervision of the passive probability average value, the cost of the active mixed mean, the passive mixed mean and the passive mean is obtained, and the second supervision cost is obtained.

[0131] In step S3011, the temperature control strategy optimization system determines the active mixed trigger prediction probability corresponding to the active sample strategy type mark, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type mark, and the passive mixed number corresponding to the passive mixed trigger prediction probability in the second number of mixed trigger prediction probabilities according to the sample strategy type mark. The sample strategy type mark is an identification of the strategy type corresponding to the strategy trigger monitoring sequence sample, which is divided into active mark and passive mark. The active mark corresponds to active strategy, such as active cooling and active heating, and the passive mark corresponds to passive strategy, such as natural heat dissipation and natural heating. The mixed trigger prediction probability is the probability value output after loading the second number of mixed sample feature vectors into the mixed trigger prediction network. The second number of mixed trigger prediction probabilities are traversed and classified according to the sample strategy type mark.

[0132] In step S3012, the temperature control strategy optimization system determines, among the first quantity of trigger prediction probabilities, an active trigger prediction probability corresponding to the active sample strategy type label, an active quantity corresponding to the active trigger prediction probability, a passive trigger prediction probability corresponding to the passive sample strategy type label, and a passive quantity corresponding to the passive trigger prediction probability, according to the sample strategy type label. The trigger prediction probability is a probability value output after loading the sample feature vector into the strategy trigger prediction network. The temperature control strategy optimization system also classifies the first quantity of trigger prediction probabilities according to the sample strategy type label.

[0133] In step S3013, the temperature control strategy optimization system obtains an active mixed mean value according to the active mixed trigger prediction probability and the active mixed quantity, and obtains a passive mixed mean value according to the passive mixed trigger prediction probability and the passive mixed quantity. The mean value is the average value of a group of data, which is used to reflect the central tendency of the data. The calculation formula of the active mixed mean value is: active mixed mean value = sum of active mixed trigger prediction probabilities / active mixed quantity. For example, the active mixed quantity is 30, and the active mixed trigger prediction probabilities are 0.2, 0.3,..., 0.4, etc. The sum of these probability values divided by 30 can obtain the active mixed mean value. Similarly, the calculation formula of the passive mixed mean value is: passive mixed mean value = sum of passive mixed trigger prediction probabilities / passive mixed quantity. The temperature control strategy optimization system can obtain the corresponding mean value by summing and then dividing by the corresponding quantity through loop iteration of the active mixed trigger prediction probability and the passive mixed trigger prediction probability.

[0134] In step S3014, the temperature control strategy optimization system obtains an active mean value according to the active trigger prediction probability and the active quantity, and obtains a passive mean value according to the passive trigger prediction probability and the passive quantity. The calculation formula of the active mean value is: active mean value = sum of active trigger prediction probabilities / active quantity. For example, the active quantity is 80, and the active trigger prediction probabilities are 0.1, 0.2,..., 0.3, etc. The sum of these probability values divided by 80 can obtain the active mean value. The calculation formula of the passive mean value is: passive mean value = sum of passive trigger prediction probabilities / passive quantity. The temperature control strategy optimization system uses a similar calculation method as step S3013 to obtain the corresponding mean value by summing and then dividing by the corresponding quantity through loop iteration of the active trigger prediction probability and the passive trigger prediction probability.

[0135] In step S3015, if the uniformity supervision requirement is the uniformity supervision of the active probability average value, the temperature control strategy optimization system obtains the cost of the active mixed average value, the passive mixed average value and the active average value to obtain the second supervision cost. The uniformity supervision of the active probability average value aims to ensure that the probability average values output by different networks (mixed trigger prediction network and strategy trigger prediction network) of the active strategy are consistent. The cost acquisition can use various loss functions, such as the mean square error loss function (MSE). In this case, the active mixed average value, the passive mixed average value and the active average value are substituted into the loss function for calculation. For example, the active mixed average value is 0.3, the passive mixed average value is 0.2, and the active average value is 0.35. Assuming that the active average value is the true value, and the active mixed average value and the passive mixed average value are the predicted values, the second supervision cost is calculated by the mean square error loss function. The temperature control strategy optimization system can adjust the network parameters of the mixed trigger prediction network according to the calculation result of the loss function to reduce the cost and improve the prediction accuracy.

[0136] In step S3016, if the uniformity supervision requirement is the uniformity supervision of the passive probability average value, the temperature control strategy optimization system obtains the cost of the active mixed average value, the passive mixed average value and the passive average value to obtain the second supervision cost. The uniformity supervision of the passive probability average value aims to ensure that the probability average values output by different networks are consistent. Similarly, the mean square error loss function is used for cost acquisition, and the active mixed average value, the passive mixed average value and the passive average value are substituted into the loss function. For example, the active mixed average value is 0.25, the passive mixed average value is 0.3, and the passive average value is 0.32. Assuming that the passive average value is the true value, and the active mixed average value and the passive mixed average value are the predicted values, the second supervision cost is calculated by the mean square error loss function. The temperature control strategy optimization system adjusts the network parameters of the mixed trigger prediction network according to the calculated second supervision cost to make the probability prediction of the passive strategy by different networks more consistent.

[0137] As another embodiment, after step S30 of loading the second number of mixed sample feature vectors into the mixed trigger prediction network and outputting the second number of mixed trigger prediction probabilities based on the mixed trigger prediction network, the method further comprises: step S30A: obtaining the actual trigger rate corresponding to each strategy trigger monitoring sequence sample; step S30B: according to the uniformity supervision requirement, obtaining the cost of the second number of mixed trigger prediction probabilities and the actual trigger rate to obtain the third supervision cost; and step S30C: adjusting the network parameters of the mixed trigger prediction network according to the third supervision cost.

[0138] In step S30A, the temperature control strategy optimization system obtains the actual triggering rate corresponding to each strategy triggering monitoring sequence sample. The actual triggering rate refers to the frequency of triggering of the strategy corresponding to the strategy triggering monitoring sequence sample in actual operation. For example, for a sample of a temperature control strategy of an electric vehicle battery, 100 monitoring times are performed in a period of time, and the corresponding cooling strategy is actually triggered 20 times. Therefore, the actual triggering rate of the sample is 20%. The temperature control strategy optimization system can collect actual operation data, count the number of times of triggering of the strategy corresponding to each strategy triggering monitoring sequence sample, and divide the total number of monitoring times, thereby obtaining the actual triggering rate.

[0139] In step S30B, according to the uniformity supervision requirement, the cost of the second number of mixed trigger prediction probabilities and the actual trigger rate is obtained, and the third supervision cost is obtained. This step aims to ensure that the mixed trigger prediction probability output by the mixed trigger prediction network is consistent or similar to the actual trigger rate in some statistics. The specific steps are as follows: First, according to the sample strategy type label, the warm control strategy optimization system determines the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability among the second number of mixed trigger prediction probabilities. For example, among the second number of 50 mixed trigger prediction probabilities, the sample strategy type label shows that 20 belong to the active sample strategy type label, so the corresponding probability value of these 20 is the active mixed trigger prediction probability, and the active mixed number is 20; the remaining 30 belong to the passive sample strategy type label, and the corresponding probability value is the passive mixed trigger prediction probability, and the passive mixed number is 30. Next, according to the sample strategy type label, the warm control strategy optimization system determines the active actual trigger rate corresponding to the active sample strategy type label, the active sample number corresponding to the active actual trigger rate, the passive actual trigger rate corresponding to the passive sample strategy type label, and the passive sample number corresponding to the passive actual trigger rate among the first number of actual trigger rates. For example, among the first number of 100 actual trigger rates, 40 correspond to the active sample strategy type label, so the corresponding actual trigger rate of these 40 is the active actual trigger rate, and the active sample number is 40; 60 correspond to the passive sample strategy type label, and the corresponding actual trigger rate is the passive actual trigger rate, and the passive sample number is 60. Then, according to the active mixed trigger prediction probability and the active mixed number, the warm control strategy optimization system obtains the active mixed mean, and the formula is: active mixed mean = sum of active mixed trigger prediction probability / active mixed number. Similarly, according to the passive mixed trigger prediction probability and the passive mixed number, the passive mixed mean is obtained, and the formula is: passive mixed mean = sum of passive mixed trigger prediction probability / passive mixed number. According to the active actual trigger rate and the active sample number, the active trigger mean is obtained, and the formula is: active trigger mean = sum of active actual trigger rate / active sample number. According to the passive actual trigger rate and the passive sample number, the passive trigger mean is obtained, and the formula is: passive trigger mean = sum of passive actual trigger rate / passive sample number. If the uniformity supervision requirement is the uniformity supervision of the active probability average value, the warm control strategy optimization system obtains the third supervision cost by performing cost acquisition on the active mixed mean, the passive mixed mean, and the active trigger mean. For example, the mean square error loss function (MSE) can be used for cost acquisition.If the uniformity supervision requirement is passive probability average value uniformity supervision, the temperature control strategy optimization system obtains the cost of active mixed average value, passive mixed average value and passive trigger average value to obtain a third supervision cost. Through these steps and calculations, the temperature control strategy optimization system can accurately obtain the cost of mixed trigger prediction probability and actual trigger rate according to the uniformity supervision requirement.

[0140] In step S30C, the temperature control strategy optimization system adjusts the network parameters of the mixed trigger prediction network according to the third supervision cost. The third supervision cost reflects the difference between the mixed trigger prediction probability output by the mixed trigger prediction network and the actual trigger rate. The goal of the temperature control strategy optimization system is to reduce this difference by adjusting the network parameters and improve the accuracy of network prediction. The temperature control strategy optimization system can use gradient descent algorithm and its variants, such as stochastic gradient descent (SGD), Adagrad, Adadelta, Adam, etc. The gradient of the network parameters is calculated according to the third supervision cost, and the network parameters are updated in the opposite direction of the gradient. In each training iteration, the gradient is calculated and the network parameters are updated until the third supervision cost converges to a small value or reaches the preset number of training times.

[0141] As an implementation, in step S30B, the cost of the second number of mixed trigger prediction probabilities and the actual trigger rates is obtained according to the uniformity supervision requirement, to obtain the third supervision cost, including: in step S30B1, according to the sample strategy type label, in the second number of mixed trigger prediction probabilities, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability are determined; in step S30B2, according to the sample strategy type label, in the first number of actual trigger rates, the active actual trigger rate corresponding to the active sample strategy type label, the active sample number corresponding to the active actual trigger rate, the passive actual trigger rate corresponding to the passive sample strategy type label, and the passive sample number corresponding to the passive actual trigger rate are determined; in step S30B3, the active mixed mean value is obtained according to the active mixed trigger prediction probability and the active mixed number, and the passive mixed mean value is obtained according to the passive mixed trigger prediction probability and the passive mixed number; in step S30B4, the active trigger mean value is obtained according to the active actual trigger rate and the active sample number, and the passive trigger mean value is obtained according to the passive actual trigger rate and the passive sample number; in step S30B5, if the uniformity supervision requirement is the uniformity supervision of the active probability average value, the cost of the active mixed mean value, the passive mixed mean value and the active trigger mean value is obtained, to obtain the third supervision cost; in step S30B6, if the uniformity supervision requirement is the uniformity supervision of the passive probability average value, the cost of the active mixed mean value, the passive mixed mean value and the passive trigger mean value is obtained, to obtain the third supervision cost.

[0142] In step S30B1, the temperature control strategy optimization system determines, according to the sample strategy type label, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability, among the second number of mixed trigger prediction probabilities. The sample strategy type label is used to distinguish different types of strategies, such as active strategies (such as active cooling, active heating, etc.) and passive strategies (such as natural cooling, natural heating, etc.). The mixed trigger prediction probability is the output result obtained after loading the second number of mixed sample feature vectors into the mixed trigger prediction network. The second number of mixed trigger prediction probabilities are traversed, and classified according to the sample strategy type label. For example, if the second number is 80, and the sample strategy type label shows that 35 of them belong to the active sample strategy type label, then the 35 corresponding probability values are the active mixed trigger prediction probability, and the active mixed number is 35; the remaining 45 belong to the passive sample strategy type label, and the corresponding probability values are the passive mixed trigger prediction probability, and the passive mixed number is 45. The implementation of this step can use data screening and statistical tools, such as using the query function of the database to screen and count the mixed trigger prediction probability according to the sample strategy type label.

[0143] In step S30B2, the temperature control strategy optimization system determines, according to the sample strategy type label, the active actual trigger rate corresponding to the active sample strategy type label, the active sample number corresponding to the active actual trigger rate, the passive actual trigger rate corresponding to the passive sample strategy type label, and the passive sample number corresponding to the passive actual trigger rate, among the first number of actual trigger rates. The actual trigger rate is the frequency of a certain strategy being triggered in an actual running scene. The temperature control strategy optimization system also classifies the first number of actual trigger rates according to the sample strategy type label. Assuming that the first number is 150, and the active sample strategy type label corresponds to 60, then the 60 corresponding actual trigger rates are the active actual trigger rate, and the active sample number is 60; the passive sample strategy type label corresponds to 90, and the corresponding actual trigger rate is the passive actual trigger rate, and the passive sample number is 90. This step can be implemented by using data classification and statistical tools to classify the actual trigger rate data according to the sample strategy type label and count the number of each category.

[0144] In step S30B3, the active mixing mean value is obtained according to the active mixing trigger prediction probability and the active mixing number, and the passive mixing mean value is obtained according to the passive mixing trigger prediction probability and the passive mixing number. The mean value is the average of a set of data, which can reflect the central tendency of the data. The calculation formula of the active mixing mean value is: active mixing mean value = sum of active mixing trigger prediction probability / active mixing number. For example, the active mixing number is 35, and the active mixing trigger prediction probabilities are 0.22, 0.25,..., 0.3, etc. The temperature control strategy optimization system adds these probability values and divides by 35 to obtain the active mixing mean value. Similarly, the calculation formula of the passive mixing mean value is: passive mixing mean value = sum of passive mixing trigger prediction probability / passive mixing number.

[0145] In step S30B4, the active trigger mean value is obtained according to the active actual trigger rate and the active sample number, and the passive trigger mean value is obtained according to the passive actual trigger rate and the passive sample number. The calculation formula of the active trigger mean value is: active trigger mean value = sum of active actual trigger rate / active sample number. For example, the active sample number is 60, and the active actual trigger rates are 0.18, 0.2,..., 0.23, etc. The temperature control strategy optimization system adds these actual trigger rates and divides by 60 to obtain the active trigger mean value. The calculation formula of the passive trigger mean value is: passive trigger mean value = sum of passive actual trigger rate / passive sample number. The temperature control strategy optimization system uses a similar method to calculate the active mixing mean value and the passive mixing mean value, that is, by iterating the active actual trigger rate and the passive actual trigger rate, summing them up and then dividing by the corresponding number to obtain the corresponding mean value.

[0146] In step S30B5, if the uniformity supervision requirement is the uniformity supervision of the active probability average value, the temperature control strategy optimization system obtains the third supervision cost by cost acquisition of the active mixing mean value, the passive mixing mean value and the active trigger mean value. The uniformity supervision of the active probability average value aims to ensure that the probability average value output by the mixed trigger prediction network of the active strategy is consistent or similar to the actual average value of the active trigger rate. The cost acquisition can use various loss functions, such as the mean square error loss function (MSE).

[0147] In step S30B6, if the uniformity supervision requirement is the uniformity supervision of the passive probability average value, the temperature control strategy optimization system obtains the third supervision cost by cost acquisition of the active mixing mean value, the passive mixing mean value and the passive trigger mean value. The purpose of the uniformity supervision of the passive probability average value is to ensure that the probability average value output by the mixed trigger prediction network of the passive strategy matches the actual average value of the passive trigger rate. Similarly, the mean square error loss function is used for cost acquisition, and the active mixing mean value, the passive mixing mean value and the passive trigger mean value are substituted into the loss function.

[0148] Please refer to Figure 2 As a structural block diagram of the temperature control strategy optimization system 1000 of the present application, the temperature control strategy optimization system 1000 comprises a computing unit 1001 which can perform various appropriate actions and processes according to a computer program stored in a ROM 1002 (i.e. read-only memory) or a computer program loaded from a storage unit 1008 to a RAM 1003 (i.e. random access memory). In the RAM 1003, various programs and data required for the operation of the temperature control strategy optimization system 1000 can also be stored. The computing unit 1001, the ROM 1002 and the RAM 1003 are connected to each other through a bus 1004. An I / O interface 1005 (i.e. input / output interface) is also connected to the bus 1004.

[0149] Various components in the temperature control strategy optimization system 1000 are connected to the I / O interface 1005, including an input unit 1006, an output unit 1007, a storage unit 1008 and a communication unit 1009. The input unit 1006 can be a device capable of inputting information to the temperature control strategy optimization system 1000, which can receive inputted digital or character information and generate key signal inputs related to user settings and / or function controls of the electronic device. The output unit 1007 can be any type of device capable of presenting information, and can include but not limited to a display, a speaker, a video / audio output terminal, a vibrator and / or a printer. The storage unit 1008 can include but not limited to a magnetic disk, an optical disk. The communication unit 1009 allows the temperature control strategy optimization system 1000 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include but not limited to a modem, a network card, an infrared communication device, a wireless communication transceiver and / or a chipset.

[0150] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include but not limited to a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1001 performs various methods and processes described above.

[0151] That is, the present application provides a temperature control strategy optimization system 1000 comprising at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the neural network driven electric vehicle temperature control strategy optimization method described above.

Claims

1. A neural network driven electric vehicle temperature control strategy optimization method, characterized in that, The method comprises: characterizing each of the first number of strategy trigger monitoring sequences corresponding to the first number of candidate electric vehicle battery temperature control strategies, to obtain a target feature vector corresponding to each of the strategy trigger monitoring sequences; determining, according to a trigger prediction reference value, a second number of target hybrid feature vectors from the first number of target feature vectors, the second number being a positive integer not greater than the first number; loading the second number of target hybrid feature vectors into a hybrid trigger prediction network respectively, and outputting a second number of pending hybrid trigger prediction probabilities based on the hybrid trigger prediction network; determining, according to a hybrid trigger prediction reference value, a third number of target hybrid trigger prediction probabilities from the second number of pending hybrid trigger prediction probabilities, wherein the third number is a positive integer not greater than the second number; weighting the third number of target hybrid trigger prediction probabilities according to the weighting coefficients of the candidate electric vehicle battery temperature control strategies corresponding to the third number of target hybrid trigger prediction probabilities respectively, and taking the candidate electric vehicle battery temperature control strategy corresponding to the maximum target hybrid trigger prediction probability after weighting as the electric vehicle temperature control strategy; wherein the training process of the hybrid trigger prediction network comprises: characterizing each of the first number of strategy trigger monitoring sequence samples, to obtain a sample feature vector corresponding to each of the strategy trigger monitoring sequence samples, wherein each of the sample feature vectors carries a corresponding sample strategy type label; determining, according to a trigger prediction reference value, a second number of hybrid sample feature vectors from the first number of sample feature vectors, wherein the second number is a positive integer not greater than the first number; loading the second number of hybrid sample feature vectors into a hybrid trigger prediction network respectively, and outputting a second number of hybrid trigger prediction probabilities based on the hybrid trigger prediction network; loading the second number of hybrid sample feature vectors into a collaborative trigger prediction network respectively, and outputting a second number of collaborative trigger prediction probabilities based on the collaborative trigger prediction network; obtaining a first supervision cost according to the uniformity supervision requirement, according to the second number of hybrid trigger prediction probabilities and the second number of collaborative trigger prediction probabilities; adjusting network parameters of the hybrid trigger prediction network according to the first supervision cost; obtaining a dispersion cost according to the dispersion supervision requirement, according to the second number of hybrid trigger prediction probabilities; the adjusting network parameters of the hybrid trigger prediction network according to the first supervision cost comprises: adjusting network parameters of the hybrid trigger prediction network according to the first supervision cost and the dispersion cost; the determining, according to a trigger prediction reference value, a second number of hybrid sample feature vectors from the first number of sample feature vectors comprises: determining a policy trigger prediction network according to the sample policy type label, and loading a sample feature vector corresponding to the sample policy type label into the policy trigger prediction network, and outputting a trigger prediction probability corresponding to the sample feature vector based on the policy trigger prediction network; selecting, from the first number of sample feature vectors, a sample feature vector with a trigger prediction probability not less than the trigger prediction reference value, to obtain a second number of mixed sample feature vectors; after loading the second number of mixed sample feature vectors into the mixed trigger prediction network and outputting a second number of mixed trigger prediction probabilities based on the mixed trigger prediction network, the method further comprises: obtaining a second supervision cost according to the second number of mixed trigger prediction probabilities and the trigger prediction probability based on the uniformity supervision requirement; adjusting network parameters of the mixed trigger prediction network according to the second supervision cost.

2. The method of claim 1, wherein, after loading the second number of mixed sample feature vectors into the mixed trigger prediction network and outputting a second number of mixed trigger prediction probabilities based on the mixed trigger prediction network, the method further comprises: obtaining an actual trigger rate corresponding to each policy trigger monitoring sequence sample; obtaining a third supervision cost according to the second number of mixed trigger prediction probabilities and the actual trigger rate based on the uniformity supervision requirement; adjusting network parameters of the mixed trigger prediction network according to the third supervision cost.

3. The method of claim 1, wherein, The policy trigger prediction network comprises an active policy trigger prediction network and a passive policy trigger prediction network, and the trigger prediction probability comprises an active trigger prediction probability and a passive trigger prediction probability. The method of determining a policy trigger prediction network according to the sample policy type label, and loading a sample feature vector corresponding to the sample policy type label into the policy trigger prediction network, and outputting a trigger prediction probability corresponding to the sample feature vector based on the policy trigger prediction network, comprises: if the sample policy type label is an active label, determining the active policy trigger prediction network, and loading a sample feature vector corresponding to the active label into the active policy trigger prediction network, and outputting an active trigger prediction probability corresponding to the sample feature vector based on the active policy trigger prediction network; if the sample policy type label is a passive label, determining the passive policy trigger prediction network, and loading a sample feature vector corresponding to the passive label into the passive policy trigger prediction network, and outputting a passive trigger prediction probability corresponding to the sample feature vector based on the passive policy trigger prediction network; obtaining an active cost according to the active label and the active trigger prediction probability, and adjusting network parameters of the active policy trigger prediction network according to the active cost; obtaining a passive cost according to the passive label and the passive trigger prediction probability, and adjusting network parameters of the passive policy trigger prediction network according to the passive cost; The sample feature vectors in the first quantity are selected to obtain the second quantity of the mixed sample feature vectors, including: The sample feature vectors in the first quantity are selected to obtain the second quantity of the mixed sample feature vectors, including:

4. The method of claim 1, wherein, If the uniformity supervision requirement is the probability average value uniformity supervision, then the cost of the second quantity of the mixed trigger prediction probability and the second quantity of the cooperative trigger prediction probability is obtained according to the uniformity supervision requirement, and the first supervision cost is obtained, including: According to the sample strategy type label, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed quantity corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed quantity corresponding to the passive mixed trigger prediction probability are determined in the second quantity of the mixed trigger prediction probability. According to the sample strategy type label, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed quantity corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed quantity corresponding to the passive mixed trigger prediction probability are determined in the second quantity of the mixed trigger prediction probability. According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. The active mixed mean value, the passive mixed mean value, the active cooperative mean value, and the passive cooperative mean value are costed to obtain the first supervision cost.

5. The method of claim 4, wherein, If the dispersion degree supervision requirement is the probability distribution dispersion degree supervision, then the cost of the second quantity of the mixed trigger prediction probability is obtained according to the dispersion degree supervision requirement, and the dispersion cost is obtained, including: According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. The active mixed mean value, the passive mixed mean value, the active cooperative mean value, and the passive cooperative mean value are costed to obtain the first supervision cost. If the dispersion degree supervision requirement is the probability distribution dispersion degree supervision, then the cost of the second quantity of the mixed trigger prediction probability is obtained according to the dispersion degree supervision requirement, and the dispersion cost is obtained, including: According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. According to the active mixed trigger prediction probability and the active mixed quantity, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed quantity, the passive mixed mean value is obtained. The active mixed mean value, the passive mixed mean value, the active cooperative mean value, and the passive cooperative mean value are costed to obtain the first supervision cost.

6. The method of claim 1, wherein, The cost is obtained according to the uniformity supervision requirement, the second number of the mixed trigger prediction probability and the trigger prediction probability, and the second supervision cost is obtained, including: According to the sample strategy type label, in the second number of the mixed trigger prediction probability, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability are determined; According to the sample strategy type label, in the first number of the trigger prediction probability, the active trigger prediction probability corresponding to the active sample strategy type label, the active number corresponding to the active trigger prediction probability, the passive trigger prediction probability corresponding to the passive sample strategy type label, and the passive number corresponding to the passive trigger prediction probability are determined; According to the active mixed trigger prediction probability and the active mixed number, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed number, the passive mixed mean value is obtained; According to the active trigger prediction probability and the active number, the active mean value is obtained, and according to the passive trigger prediction probability and the passive number, the passive mean value is obtained; If the uniformity supervision requirement is the uniformity supervision of the active probability average value, the active mixed mean value, the passive mixed mean value and the active mean value are costed, and the second supervision cost is obtained; If the uniformity supervision requirement is the uniformity supervision of the passive probability average value, the active mixed mean value, the passive mixed mean value and the passive mean value are costed, and the second supervision cost is obtained.

7. The method of claim 2, wherein, The cost is obtained according to the uniformity supervision requirement, the second number of the mixed trigger prediction probability and the actual trigger rate, and the third supervision cost is obtained, including: According to the sample strategy type label, in the second number of the mixed trigger prediction probability, the active mixed trigger prediction probability corresponding to the active sample strategy type label, the active mixed number corresponding to the active mixed trigger prediction probability, the passive mixed trigger prediction probability corresponding to the passive sample strategy type label, and the passive mixed number corresponding to the passive mixed trigger prediction probability are determined; According to the sample strategy type label, in the first number of the actual trigger rate, the active actual trigger rate corresponding to the active sample strategy type label, the active sample number corresponding to the active actual trigger rate, the passive actual trigger rate corresponding to the passive sample strategy type label, and the passive sample number corresponding to the passive actual trigger rate are determined; According to the active mixed trigger prediction probability and the active mixed number, the active mixed mean value is obtained, and according to the passive mixed trigger prediction probability and the passive mixed number, the passive mixed mean value is obtained; According to the active actual trigger rate and the active sample number, the active trigger mean value is obtained, and according to the passive actual trigger rate and the passive sample number, the passive trigger mean value is obtained; If the uniformity supervision requirement is active probability average value uniformity supervision, cost acquisition is performed on the active mixed average value, the passive mixed average value and the active triggered average value to obtain the third supervision cost; If the uniformity supervision requirement is passive probability average value uniformity supervision, cost acquisition is performed on the active mixed average value, the passive mixed average value and the passive triggered average value to obtain the third supervision cost.

8. A temperature control strategy optimization system, characterized by, The method comprises the following steps: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

Images