Energy system life prediction method and system based on interpretable deep learning

Abstract

The invention discloses an interpretable deep learning-based energy system life prediction method and system. According to the method, to-be-tested data of an energy system is input into a CNN-BiLSTM space-time double-flow model, and a life prediction result is obtained; the CNN-BiLSTM space-time double-flow model comprises a depth feature extraction module and an interpretability module, the depth feature extraction module comprises a CNN feature extraction module, a BiLSTM feature extraction module and a full-connection layer, the CNN feature extraction module is used for extracting space features, the BiLSTM feature extraction module is used for extracting time features, and the full-connection layer is used for extracting the time features; the full connection layer is used for predicting the time features and the space features after splicing and fusion and outputting a life prediction result; and the interpretability module is used for visualizing the output of the CNN feature extraction module. According to the method, the time-space information of energy system degradation can be captured, the model feature extraction capability is improved, the prediction precision is high, and the model has interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of predictive maintenance of energy systems, and in particular to an energy system life prediction method and system based on explainable deep learning. Background Art

[0002] Electric energy is becoming more and more popular as a new type of environmentally friendly energy, and lithium-ion batteries are one of the most commonly used rechargeable batteries used in various mobile devices and electric vehicles. The stable operation of lithium battery packs is mainly monitored and managed by battery management systems (BMS). However, there are still some problems in the field of battery health management, such as low battery utilization leading to resource waste and inaccurate battery health prediction leading to battery explosion. SOH (state of health) prediction of lithium-ion batteries refers to determining the health state of the battery and predicting it through real-time monitoring and analysis. SOH prediction is usually achieved through a series of data analysis and deep learning algorithms, including monitoring and analyzing battery parameters such as voltage, current, temperature, capacity, and internal resistance. Predicting the SOH of lithium-ion batteries can better help manage and optimize the use and maintenance of batteries and improve utilization. In addition, by monitoring battery health and predicting the remaining service life, more reasonable charging and discharging strategies can be formulated to extend battery life. By predicting the SOH of the battery, aging and safety issues can be discovered in time, the safety of the battery can be improved, safety risks and unnecessary maintenance and replacement can be avoided, and the overall cost can be reduced.

[0003] At present, the most commonly used SOH prediction methods for lithium-ion batteries can be divided into empirical model method, filter prediction method and data-driven method. The empirical model method is to construct the mathematical relationship between the number of battery charge and discharge cycles and the current capacity through expert experience, and select the mathematical model with the smallest fitting error as the SOH prediction model of lithium batteries. The filter prediction method refers to updating and correcting the battery capacity based on various filtering algorithms such as Kalman filtering, which is usually combined with the empirical model method. The data-driven method is completely different from the first two methods. The data-driven method does not rely on constructing physical and mathematical models of lithium battery degradation, nor does it need to pay too much attention to the internal and external characteristics of lithium-ion battery charging and discharging. Through big data and machine learning technology, the complex nonlinear mapping relationship between the monitored working parameters and the battery health status can be learned. This method has high prediction accuracy and has become a new research hotspot.

[0004] Many new methods have been developed for SOH and RUL (remaining useful life) prediction of lithium-ion batteries, and some results have been achieved. However, existing deep neural networks for lithium-ion battery SOH prediction have some problems, such as lack of strong generalization ability under variable conditions and weak interpretation of the model during the decision-making process. Most deep neural networks in the field of lithium-ion battery SOH prediction are black box models, which makes it difficult to understand the deep working principle of the neural network and why the network predicts the corresponding value.

[0005] In the field of lithium-ion battery SOH prediction, existing models are deficient in feature extraction. For example, the technical solution disclosed by Yukai Tian et al. in "State-of-Health Prediction of Lithium-Ion Batteries Based on CNN-BiLSTM-AM" does not have a good feature extraction module designed for lithium-ion batteries. Existing single data-driven models or end-to-end models are difficult to take into account the multi-faceted feature extraction and accurate SOH prediction of lithium-ion batteries. For example, when only simple CNN (convolutional neural network) or LSTM (long short-term memory network) is used, it may not be possible to fully mine the complex features in the battery time series data, and the features reflected by the complex electrochemical changes inside the battery are not fully captured, resulting in limited prediction accuracy.

[0006] Most existing models lack interpretability. In practical applications, it is difficult for users to understand the basis and process of the model's prediction. For example, when some deep learning models make SOH predictions, they only output the prediction results, but the decision-making mechanism within the model cannot clearly present why the result is obtained. This makes it difficult for technicians to determine the problem in the battery management system when the prediction results are biased or the model needs to be optimized, which increases the difficulty of model application and improvement. Summary of the invention

[0007] In order to solve at least one of the problems existing in the prior art, the present invention provides an energy system life prediction method and system based on interpretable deep learning, and uses the feature extraction modules of CNN and BILSTM to extract spatial features and temporal features respectively to improve the feature learning ability of the neural network; a spatiotemporal dual-stream structure is used to deeply mine and collaboratively integrate the extracted features, so as to more accurately and comprehensively capture the complex change laws of the energy system in different spatiotemporal dimensions, improve the prediction performance of the energy system life, and introduce 1D Grad-CAM++ feature visualization technology to realize the visualization of specific areas of interest of the input data and improve the interpretability of the model.

[0008] In order to achieve the purpose of the present invention, the energy system life prediction method provided by the present invention is to input the test data of the energy system into the trained CNN-BiLSTM spatiotemporal dual-stream model to obtain the life prediction result of the energy system; The CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module. The deep feature extraction module includes a CNN feature extraction module, a BiLSTM feature extraction module and a fully connected layer. The CNN feature extraction module is used to extract the spatial features of the data to be tested, and the BiLSTM feature extraction module is used to extract the temporal features of the data to be tested. The fully connected layer is used to predict the spliced ​​and fused temporal features and spatial features, and output the life prediction results of the energy system. The interpretability module visualizes the output of the CNN feature extraction module through 1D Grad-CAM++ feature visualization technology, and maps the output of the CNN feature extraction module back to the dimension of the data to be tested.

[0009] Furthermore, the CNN feature extraction module includes a convolution layer, an activation function, and an average pooling layer. The formula of the CNN feature extraction module is as follows:

[0010] in, Represents the abstract spatial features extracted by the CNN feature extraction module, Indicates the data to be tested. represents the activation function, represents the weight, Indicates deviation.

[0011] Furthermore, the BiLSTM feature extraction module includes a BiLSTM layer and a Flatten layer. The BiLSTM layer includes multiple LSTM layers. Each LSTM layer includes a forward LSTM and a reverse LSTM. The BiLSTM layer outputs a time feature extracted from the data to be tested. The time feature is a three-dimensional tensor. The Flatten layer is used to convert the three-dimensional tensor into a two-dimensional tensor.

[0012] Furthermore, the formula of 1D Grad-CAM++ feature visualization technology is as follows:

[0013]

[0014] in, is the weight of the CNN feature extraction module, indicating the The value of the feature for predicting the final life span The weight of It is The weight of each feature; It is the output of the CNN feature extraction module, which refers to the input feature In the Channels and The value of the position; is the Grad-CAM++ heat map; It refers to The value of a feature.

[0015] Furthermore, the energy system includes a lithium-ion battery or an insulated gate bipolar transistor.

[0016] Furthermore, when predicting a lithium-ion battery, the data to be measured in the energy system is a lithium-ion battery capacity degradation sequence that is sensitive to battery capacity and is obtained based on voltage, current and temperature data over the entire life cycle of the lithium-ion battery.

[0017] Furthermore, during training, the lithium-ion battery capacity degradation sequence is obtained in the following manner: A lithium-ion battery charge and discharge data set is acquired or constructed, wherein the lithium-ion battery charge and discharge data set includes voltage, current and temperature data of the lithium-ion battery over its entire life cycle, and physical features sensitive to battery capacity are extracted based on the lithium-ion battery charge and discharge data set to obtain a lithium-ion battery capacity degradation sequence.

[0018] Furthermore, when predicting an insulated gate bipolar transistor, the data to be measured of the energy system is the degradation data of voltage and current of the insulated gate bipolar transistor over its entire life cycle.

[0019] Preferably, the steps of the method of the present invention for predicting the life of a lithium-ion battery can be summarized as follows: Step 1: Model training. The model is trained using a lithium-ion battery charge and discharge dataset, the raw data of which contains voltage, current, and temperature data over the entire life cycle of the lithium-ion battery.

[0020] Step 2: Data preprocessing. Extract features sensitive to battery capacity from the voltage, current and temperature signals in the raw data, such as the 7 features mentioned, which are the time required for the charging voltage to rise to above 4.2V, the time required for the discharge voltage to drop below 3.2V, the amount of charging current drop at equal time intervals, the amount of discharge voltage drop at equal time intervals, the time required for the discharge voltage to reach the minimum value, the discharge power, and the time when the temperature reaches the peak during the discharge process, and normalize them.

[0021] Step 3: Input the preprocessed data into the constructed CNN-BiLSTM spatiotemporal dual-stream model and train it.

[0022] Step 4: Model testing. Extract features sensitive to capacitance from the voltage, current, and temperature signals in the original data, and then input them into the trained CNN-BiLSTM spatiotemporal dual-stream model to obtain the SOH prediction results.

[0023] Step 5: Life test: The voltage, current and temperature data of the lithium-ion battery to be tested over its entire life cycle are preprocessed as in step 2 above, and then input into the trained CNN-BiLSTM spatiotemporal dual-stream model to obtain the life of the lithium-ion battery.

[0024] Step 6: Visualization. Use 1D Grad-CAM++ to visualize the importance of features for life prediction at different life cycles.

[0025] The energy system life prediction system based on explainable deep learning provided by the present invention is used to implement the above method, and the system includes the following units: A detection unit is used to use a CNN-BiLSTM spatiotemporal dual-stream model to predict the life of the energy system; A model unit is used for training to obtain a CNN-BiLSTM spatiotemporal dual-stream model, wherein the CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module, wherein the deep feature extraction module includes a CNN feature extraction module, a BiLSTM feature extraction module and a fully connected layer, wherein the CNN feature extraction module is used to extract spatial features of the energy system test data, the BiLSTM feature extraction module is used to extract temporal features of the energy system test data, and the fully connected layer is used to predict the spliced ​​and fused temporal features and spatial features, and output the prediction results of the energy system; The interpretability module is used to visualize the output of the CNN feature extraction module through the 1D Grad-CAM++ feature visualization technology, map the output of the CNN feature extraction module back to the dimension of the energy system data to be tested, and output a class activation map reflecting the importance of the feature.

[0026] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the energy system life prediction method based on explainable deep learning when executing the computer program.

[0027] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the energy system life prediction method based on explainable deep learning.

[0028] Compared with the prior art, the present invention has the following advantages and effects: (1) The method of the present invention constructs a feature extraction technology specifically for energy system life prediction, and constructs a CNN feature extraction module and a BILSTM feature extraction module, which are respectively used to extract the spatial features and temporal features of the energy system capacity degradation sequence to improve the feature learning ability of the neural network. This deep feature extraction module is particularly suitable for situations where the sample noise is large. Since energy systems, especially lithium-ion batteries, are faced with various environmental noises and changes in operating conditions in actual use, the CNN feature extraction module can effectively extract useful spatial features and temporal features from these noises, providing a solid foundation for subsequent life prediction.

[0029] (2) The spatiotemporal dual-stream structure is used to deeply mine and coordinately integrate the extracted features, so as to more accurately and comprehensively capture the complex changes of the energy system in different spatiotemporal dimensions and improve the prediction performance of the energy system life. By combining the two, the spatiotemporal dual-stream structure can simultaneously utilize information in the time and space dimensions, avoiding the limitations of single-dimensional information, enabling the CNN-BILSTM spatiotemporal dual-stream model to comprehensively consider the state of the energy system from different angles, providing a more comprehensive information basis for subsequent predictions, and thus more accurately predicting the life of the energy system. In actual use, the energy system will face various complex working conditions. Comprehensive consideration of time and space information enables the model to make more accurate predictions based on different working condition characteristics.

[0030] (3) The present invention can improve the interpretability of neural networks. By introducing 1D Grad-CAM++ feature visualization technology to visualize the specific mechanism of the model in predicting the life of the energy system, the importance of each feature at different time steps can be seen, so as to analyze which feature plays a greater role in different stages of the energy system life and which feature is a more important consideration in which time period. At the same time, the physical mechanism of energy system degradation can be combined to understand the degradation process of the entire life cycle, so as to guide the model to further improve the prediction performance and select more important features for prediction at different time steps. This technology can be used to determine which time points or data segments are most critical to the model's decision to predict the life of the energy system. The visualization results can help evaluate whether the architecture of the CNN feature extraction module is reasonable. If it is found that the features that the model mainly focuses on are not the features that are truly related to the battery life, it may be necessary to adjust the structure of the CNN feature extraction module, such as adjusting the convolution kernel size, the number of layers, etc., so that the model can better focus on the features that have an important impact on the life of the energy system, rather than insignificant details or noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a lithium-ion battery life prediction method based on an interpretable neural network of CNN and BILSTM of the present invention.

[0032] Figure 2 It is a degradation curve diagram of each battery in the data set of the verification method of the present invention.

[0033] Figure 3 and Figure 4 It is a schematic diagram of the verification effect of two different batteries on the data set of the present invention.

[0034] Figure 5 This is an effect diagram of feature visualization using 1D Grad-CAM++ technology in an embodiment of the present invention.

[0035] Figure 6 It is a schematic diagram of the life prediction result of the insulated gate bipolar transistor in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0037] See also Figure 1 The energy system life prediction method based on explainable deep learning provided in this embodiment, wherein the energy system is a lithium-ion battery, comprises the following steps: Step S1: extracting physical features sensitive to battery capacity based on a lithium-ion battery charge and discharge data set to obtain a lithium-ion battery capacity degradation sequence.

[0038] In some embodiments of the present invention, according to the lithium-ion battery life prediction task, the voltage, current, and temperature data of the lithium-ion battery in the open source NASA lithium-ion battery charge and discharge data set over the entire life cycle are artificially preprocessed to extract physical features sensitive to battery capacity, and the extracted physical features are normalized by Min-max normalization (0-1 normalization), and the preprocessed physical features are subsequently input into the feature extraction module for feature extraction. Among them, normalization converts all input data into the same numerical interval, which can speed up the convergence speed of training the CNN-BILSTM spatiotemporal dual-stream model and reduce the sensitivity of the CNN-BILSTM spatiotemporal dual-stream model to data with a large numerical range. In other embodiments, the open source NASA lithium-ion battery charge and discharge data set can be replaced by a self-built data set, which is obtained by collecting the voltage, current, and temperature data of the lithium-ion battery over the entire life cycle.

[0039] In some embodiments of the present invention, seven physical features sensitive to battery capacity are extracted, namely, the time required for the charging voltage to rise to above 4.2V, the time required for the discharging voltage to drop below 3.2V, the amount of charging current drop at equal time intervals, the amount of discharging voltage drop at equal time intervals, the time required for the discharging voltage to reach the minimum value, the discharging power, and the time when the temperature during the discharging process reaches the peak value. In other embodiments, the number of extracted features sensitive to battery capacity may also be other values, determined by empirical knowledge.

[0040] The present invention predicts the life of lithium-ion batteries through a CNN-BiLSTM spatiotemporal dual-stream model. The CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module, and the deep feature extraction module includes a CNN feature extraction module, a BiLSTM feature extraction module and a fully connected layer. The interpretability module visualizes the output of the CNN feature extraction module through 1D Grad-CAM++ feature visualization technology, maps the output of the CNN feature extraction module back to the dimension of the lithium-ion battery capacity degradation sequence, and outputs a class activation map reflecting the importance of the feature.

[0041] Step S2: extracting spatial high-dimensional features in the lithium-ion battery capacity degradation sequence through a CNN feature extraction module.

[0042] In order to predict the life of lithium-ion batteries, the present invention provides a CNN (convolutional neural network) feature extraction module to extract the spatial features of the lithium-ion battery capacity degradation sequence.

[0043] The CNN feature extraction module is particularly suitable for situations where the sample noise is large. CNN can be used for automatic feature extraction and dimensionality reduction to reduce the workload of data preprocessing. Since lithium-ion batteries are faced with various environmental noises and changes in operating conditions in actual use, the CNN feature extraction module can effectively extract useful spatial features from these noises, providing a solid foundation for subsequent life prediction.

[0044] The CNN feature extraction module includes a convolutional layer and a pooling layer, wherein the convolutional layer is used to extract the spatial features of the lithium-ion battery capacity degradation sequence, and the pooling layer is responsible for reducing the feature dimension to enhance the generalization ability of the CNN feature extraction module.

[0045] A typical convolutional neural network includes multiple alternating convolutional layers and pooling layers. The CNN feature extraction module of the present invention includes a convolutional layer + activation function + average pooling layer. The convolutional layer is a feature extractor for extracting high-dimensional abstract features, while the main function of the pooling layer is feature dimensionality reduction. Common pooling methods include maximum pooling and average pooling. The present invention adopts average pooling. The average pooling layer is used to reduce the number of parameters and prevent overfitting while retaining the most important feature information.

[0046] The formula of the CNN feature extraction module is as follows:

[0047] in, Represents the abstract spatial features extracted by the CNN feature extraction module, represents the data to be tested, which represents the physical characteristics sensitive to the battery capacity, that is, the capacity degradation sequence of the lithium-ion battery (in some embodiments of the present invention, That is, the seven physical characteristics selected that are sensitive to battery capacity). represents the activation function, represents the weight, Indicates deviation.

[0048] In some embodiments of the present invention, there are 64 convolution kernels in the CNN feature extraction module, and the size of each convolution kernel is 3x3. Selecting 64 convolution kernels can maintain computational efficiency while capturing local features. The size of the convolution kernel is set to 3x3. The life of lithium-ion batteries is affected by many factors, and its data has rich feature dimensions. Setting 64 convolution kernels allows the CNN feature extraction module to learn different feature representations. Each convolution kernel performs a convolution operation on the input data, i.e., the capacity degradation sequence of the lithium-ion battery, to generate a feature map, so 64 convolution kernels will generate 64 different feature maps. These different feature maps can capture the features of lithium-ion battery data at different angles. For example, some convolution kernels may focus on capturing the fluctuation characteristics of the voltage curve, some may pay more attention to the periodic characteristics of the current curve, and some may be more sensitive to the characteristics of temperature changes. In this way, the CNN feature extraction module can understand and represent the state of the lithium-ion battery from multiple aspects, providing more comprehensive and diverse feature information for subsequent life prediction. The life-related data of lithium-ion batteries usually contains complex time series information, such as the changes of physical quantities such as voltage, current, and temperature over time. These data often have important correlations and patterns in the local range. The 3x3 convolution kernel can focus on the local area well and extract local features. For example, in the voltage time series data, the voltage change characteristics between several adjacent time points may have key information for predicting the aging and life of the battery. The 3x3 convolution kernel can effectively capture these local details, such as the rising or falling trend and fluctuation of the voltage in a short period of time. This local feature extraction capability helps the model learn subtle but important local features in lithium-ion battery data. These features may be the embodiment of the chemical reactions and physical processes inside the battery in the local time window, which is crucial to the degradation of battery performance and the impact on life. The 3x3 convolution kernel has a moderate receptive field, which is neither too limited to local tiny features nor too broad to ignore details. In other embodiments, the number and size of the convolution kernels can also be set to other values ​​as needed.

[0049] The activation function in the CNN feature extraction module uses the ReLU activation function. The relationship between the life of lithium-ion batteries and various factors (such as voltage, current, temperature, number of charge and discharge times, etc.) is usually nonlinear. The ReLU activation function can map the input features to a nonlinear space, allowing the neural network to better fit these nonlinear relationships. For example, the aging process of lithium-ion batteries may not change linearly with time or number of uses, and the degradation of their performance may accelerate in some stages and be relatively stable in other stages. The ReLU activation function can help CNN capture these complex nonlinear features.

[0050] The pooling layer in the CNN feature extraction module uses average pooling. The average pooling operation can process the local area of ​​the input and extract the average features in the area. In lithium-ion battery data, the features in the local area may have similarities. For example, in a short time range, the battery's voltage, current and other parameters may show certain average characteristics. Average pooling can effectively capture these local features and integrate them into a more representative feature value.

[0051] The present invention extracts physical features sensitive to battery capacity from the original data, and then uses a CNN feature extraction module to extract higher and more abstract features, namely spatial features, from the physical features sensitive to battery capacity, thereby capturing the relationship between the physical features and providing a basis for subsequent battery SOH prediction.

[0052] Step S3: Use the BiLSTM feature extraction module to extract the time high-dimensional features in the lithium-ion battery capacity degradation sequence.

[0053] A BiLSTM (bidirectional long short-term memory network) feature extraction module is constructed to extract the temporal characteristics of the lithium-ion battery capacity degradation sequence.

[0054] The BiLSTM feature extraction module includes a BiLSTM layer and a Flatten layer. The BiLSTM layer includes multiple LSTM layers. Each LSTM layer includes a forward LSTM and a reverse LSTM. The output of the BiLSTM layer is a high-dimensional time abstract feature extracted from the lithium-ion battery capacity degradation sequence. It is a three-dimensional tensor. The first dimension is the sequence length, the second dimension is the batch size, and the third dimension is the hidden layer size. The Flatten layer is connected after the BiLSTM layer. The three-dimensional tensor is converted into a two-dimensional tensor through the Flatten layer so that it can be spliced ​​with the output of the CNN feature extraction module and subsequently input into the fully connected layer for prediction.

[0055] For lithium-ion battery life prediction, the present invention can consider past information and future information at the same time through the bidirectional structure of BiLSTM in the BiLSTM layer, and can process time series data and capture temporal dependencies, which is particularly important for predicting the health of the battery because the current state of the battery is often affected by previous usage history and expected usage patterns. The BiLSTM feature extraction module of the present invention is particularly suitable for online learning scenarios. For tasks that require multi-step predictions based on historical data, such as battery life prediction, the BiLSTM feature extraction module can use its memory capacity to capture long-term dependencies. In an online learning environment, the real-time data of the battery needs to be processed and analyzed quickly, and the BiLSTM feature extraction module can adapt to such dynamic changes and update the battery health prediction in real time.

[0056] BiLSTM is a variant of recurrent neural network (RNN) and can be used to predict time series data in the forward and backward directions. LSTM is the basis of BiLSTM, and the BiLSTM feature extraction module of the present invention includes LSTM units. LSTM units have three special gate units: input gate, forget gate and output gate, which can effectively control the input, forgetting and output of information. In the LSTM unit, each time step has a hidden state and a cell state, and the formula is as follows:

[0057] in, It is the output of the input gate, which determines how much of the current input information will be written into the cell state; and Represents the sigmoid function and the hyperbolic tangent function; and Respectively represent weights and biases, which are parameters of LSTM and are obtained through training and learning. , , , is the weight matrix of the input feature to the LSTM unit, corresponding to the input gate, forget gate, unit state update and output gate, respectively. represents the input gate, represents the forget gate, Indicates that the unit status is updated, represents the output gate; , , , is the weight matrix from the hidden state of the previous time step to the current time step, corresponding to the input gate, forget gate, unit state update, and output gate; , , , , are the bias vectors corresponding to the input gate, forget gate, cell state update, and output gate; is the input feature vector of the current time step, i.e., the artificially extracted physical features sensitive to battery capacity, i.e., the capacity degradation sequence of lithium-ion batteries; is the hidden state vector of the previous time step, which represents the information of the previous time step; is the cell state vector at the previous time step, which is also the information of the previous time step; is the output of the forget gate, which determines how much of the cell state at the previous time step will be retained; is the candidate cell state, which is a function of the current input and the previous hidden state, processed by the tanh function; is the cell state at the current time step, which is a combination of the previous cell state and the newly written information; is the output of the output gate, which determines how much of the current cell state will be output; is the hidden state at the current time step, which is a function of the output gate and the cell state and serves as the input or final output of the next LSTM cell.

[0058] BiLSTM can capture the dependency between the forward and reverse parts of the input sequence (i.e., the lithium-ion battery capacity degradation sequence), process from the beginning of the input sequence to the end to extract the forward dependency features of the input sequence, and process from the end of the input sequence to the beginning to extract the backward dependency features of the input sequence, and perform better than traditional LSTM in many tasks. In some embodiments of the present invention, for energy system life prediction, in the BiLSTM feature extraction module: Hidden layer size: The size of the hidden layer is 14, which is the dimension size of the internal state in the LSTM layer, representing the dimension of the feature representation learned by the network from the input sequence. Setting the hidden layer size to 14 allows the network to learn more complex feature representations, and can perform nonlinear transformations and abstractions on the multi-dimensional features of the input (in some embodiments of the present invention, 7 physical features sensitive to battery capacity are extracted, so here are 7-dimensional features), which helps capture nonlinear relationships and complex patterns that may exist in battery life prediction. For example, simple physical features of the input (such as voltage, current, etc.) may be converted into more representative and abstract hidden states, thereby better reflecting the performance degradation process of the battery.

[0059] Number of stacked layers: The number of LSTM layers is 2. Each LSTM layer includes a forward LSTM and a reverse LSTM. Multiple LSTM layers can learn time-dependent information at different levels. For lithium-ion battery life prediction, a single LSTM layer may only learn short-term time dependencies, while multiple LSTM layers can capture longer time dependencies. For example, the first LSTM layer can capture the performance changes of the battery in a short period of time, such as the immediate reaction during the charging and discharging process. The second LSTM layer can learn longer-term trends based on the first LSTM layer, such as the performance degradation of the battery over days or weeks, which helps to more comprehensively consider the time series information related to battery life.

[0060] Bidirectional structure. Includes forward LSTM and reverse LSTM. For time series data of energy system life prediction, it is very important to consider bidirectional information. For example, in the forward LSTM, you may see the trend of battery performance degradation over time, while in the reverse LSTM, you may see the performance changes of the battery at different time points relative to the final state. Bidirectional LSTM can combine these two aspects of information, learn the performance characteristics of the battery from different angles, and improve the accuracy of battery life prediction. The fusion of forward and reverse information can provide a more comprehensive feature representation, avoiding the one-sidedness brought by considering only one direction of information, which is especially important for capturing the subtle characteristics and complex trends of battery performance changes.

[0061] The BiLSTM feature extraction module is particularly suitable for online learning scenarios. For tasks that require multi-step predictions based on historical data, such as battery life prediction, traffic flow prediction, and power demand prediction, the BiLSTM feature extraction module can use its memory capacity to capture long-term dependencies. In an online learning environment, real-time battery data needs to be processed and analyzed quickly. The BiLSTM feature extraction module can adapt to this dynamic change and update the battery health status prediction in real time.

[0062] The present invention is based on CNN and BILSTM feature extraction modules to respectively extract the spatial features and temporal features of the lithium-ion battery capacity degradation sequence to improve the feature learning ability of the neural network.

[0063] The present invention adopts the CNN-BILSTM spatiotemporal dual-stream model with a spatiotemporal dual-stream structure to deeply mine and collaboratively integrate the extracted time features and spatial features, so as to more accurately and comprehensively capture the complex change laws of lithium-ion batteries in different spatiotemporal dimensions, and improve the prediction performance of battery SOH. Using a spatiotemporal dual-stream parallel structure, the spatial feature vector extracted by the CNN feature extraction module and the time feature vector extracted by the BILSTM feature extraction module are spliced ​​to obtain a fused feature vector, and the fused feature vector is input into multiple fully connected layers to output the final life prediction result. The function of the fully connected layer is to map the fused features to the final prediction result, and the activation function used is ReLU.

[0064] Time dimension: The performance of lithium-ion batteries (such as state of charge, health status, etc.) will change over time, and these changes are affected by factors such as charge and discharge cycles, usage time, and storage time. The BILSTM feature extraction module can process time series data and capture long-term and short-term dependencies in the time dimension, such as the performance changes, aging trends, and time laws of performance degradation of batteries under different charge and discharge cycles. It can effectively encode the historical information of lithium-ion batteries and extract important features in the time series.

[0065] Spatial dimension: Lithium-ion battery is a physical system with spatial distribution. The performance and state of components such as electrodes and electrolytes at different locations inside the battery may be different, which will affect the overall performance and life. The CNN feature extraction module can effectively process the battery data in the spatial dimension, such as extracting spatial local features from sensor data at different locations of the battery (such as temperature distribution and voltage distribution in different regions) and identifying spatial structures and patterns.

[0066] Synergistic effect: By combining the two, the spatiotemporal dual-stream structure can simultaneously utilize information in the time and space dimensions, avoiding the limitations of single-dimensional information, and enabling the CNN-BILSTM spatiotemporal dual-stream model to comprehensively consider the state of the battery from different angles, providing a more comprehensive information basis for subsequent predictions, and thus more accurately predicting the battery life. In actual use, lithium-ion batteries will face a variety of complex working conditions, including different charging and discharging rates, ambient temperatures, and usage environments. These factors will affect the performance of the battery in both time and space. The spatiotemporal dual-stream structure can enable the CNN-BILSTM spatiotemporal dual-stream model to better adapt to these complex situations, because it comprehensively considers time and space information, allowing the CNN-BILSTM spatiotemporal dual-stream model to make more accurate predictions based on different working condition characteristics, rather than relying solely on single-dimensional information.

[0067] Step S4, feature visualization technology, the interpretability module uses 1D Grad-CAM++ feature visualization technology to visualize the area of ​​interest, visualizes the output of the CNN feature extraction module, maps the output of the CNN feature extraction module back to the dimension of the lithium-ion battery capacity degradation sequence, and outputs a class activation map reflecting the importance of the feature, thereby realizing visualization and feature interpretability of specific areas of interest of the input data.

[0068] The explainability module introduces 1D Grad-CAM++ feature visualization technology to determine which time points or data segments are most critical to the model's decision to predict battery life.

[0069] In the prediction of lithium-ion battery life, the features extracted by the CNN feature extraction module are usually abstract, and it is difficult to intuitively understand the contribution of these features to the final prediction results. The present invention can map these abstract features back to the dimension of the input data, i.e., the capacity degradation sequence of the lithium-ion battery, through the 1D Grad-CAM++ technology, so as to reveal which parts of the input data have the greatest impact on the prediction results, and realize the visualization and feature interpretability of specific areas of interest of the input data. For example, for the input lithium-ion battery time series data (such as voltage, current, temperature, etc.), the 1D Grad-CAM++ technology can be used to determine which time points or data segments are most critical to the decision-making of the CNN-BILSTM spatiotemporal dual-stream model to predict the battery life. By visualizing the abstract features extracted by the CNN feature extraction module, it can be checked whether there are some data areas that are mistakenly given a higher importance by the model, which may be caused by problems with data preprocessing or outliers in the data itself. For example, if the visualization results show that the model mainly focuses on an abnormal current fluctuation, and this fluctuation is caused by measurement errors, the data processing method can be improved in a targeted manner to improve the performance of the CNN-BILSTM spatiotemporal dual-stream model. The visualization results can help evaluate whether the architecture of the CNN-BILSTM spatiotemporal dual-stream model is reasonable. If it is found that the features that the CNN-BILSTM spatiotemporal dual-stream model focuses on are not the features that are truly relevant to battery life, the structure of the CNN feature extraction module can be adjusted, such as adjusting the convolution kernel size and number of layers, so that the CNN-BILSTM can better focus on features that have an important impact on battery life, rather than insignificant details or noise.

[0070] 1D Grad-CAM++ feature visualization technology is an interpretable one-dimensional signal method that generates a heat map (class activation map) to display the area in the input data (test data) that has the greatest impact on the prediction results of the CNN-BILSTM spatiotemporal dual-stream model. Grad CAM evolved from Cam. Traditional Cam requires modifying the network structure and retraining, while Grad CAM does not have such troublesome steps and is more generally applicable. 1D Grad-CAM++ uses gradient-weighted class activation heat maps and spatial attention mechanisms to enhance the discernibility of feature areas.

[0071] The formula of 1D Grad-CAM++ feature visualization technology is as follows:

[0072]

[0073] in, is the weight of the CNN feature extraction module, indicating the The value of the feature for predicting the final life span The weight of indicates the contribution of this feature to the prediction results of the CNN-BILSTM spatiotemporal dual-stream model; is the final lifespan prediction value, which is the output of the CNN-BILSTM spatiotemporal dual-stream model; It is The weight of the feature reflects the Channels and The degree of influence of the features of each position on the output of the CNN-BILSTM spatiotemporal dual-stream model; It is the output of the CNN feature extraction module, which refers to the input feature In the Channels and The values ​​at the positions are data on changes of physical characteristics sensitive to battery capacity over time (in some embodiments of the present invention, data on changes of seven manually extracted physical characteristics over time); is a Grad-CAM++ heat map, which is used to visualize the importance of different features for prediction results at different time steps; It refers to The value of a feature.

[0074] The calculation process of the 1D Grad-CAM++ formula is as follows: Calculate the gradient: First calculate the final life prediction value Relative to each feature Gradient , which indicates the degree of influence of each feature on the model's prediction results.

[0075] Global average pooling: Perform global average pooling on the gradient to obtain the weight of each feature .

[0076] Weighted summation: Use the ReLU function to weight the gradient and get the contribution weight of each feature. .

[0077] Generating heatmap: weight Apply to Features , generating the final heat map , highlighting the feature regions that have the greatest impact on the prediction results of the CNN-BILSTM spatiotemporal two-stream model.

[0078] 1D Grad-CAM++ technology is suitable for scenarios that require model interpretation, such as battery health management and safety monitoring. Through visualization technology, engineers and researchers can better understand the prediction results of the CNN-BILSTM spatiotemporal dual-stream model, so that they can intervene when necessary to improve the safety and efficiency of battery use.

[0079] In order to improve the interpretability of the CNN-BILSTM spatiotemporal dual-stream model, the present invention introduces 1D Grad-CAM++ feature visualization technology to visualize the feature importance of different time steps. For lithium-ion battery life prediction, through visualization technology, engineers and researchers can better understand the prediction results of the CNN-BILSTM spatiotemporal dual-stream model, so as to intervene when necessary and improve the safety and efficiency of battery use.

[0080] Step S5: Test the CNN-BILSTM spatiotemporal dual-stream model.

[0081] In some embodiments of the present invention, the results are verified on the NASA PCoE dataset. The experimental results show that the proposed method can effectively predict the SOH of lithium-ion batteries, and the visualization results can help better understand the CNN-BILSTM spatiotemporal dual-stream model.

[0082] In this embodiment, the method of the present invention is verified using the NASA lithium-ion battery data set collected by the NASA Ames Performance Excellence Prediction Center (PCoE). The lithium-ion battery completes charging, discharging, and impedance operations at room temperature (24°C). First, there is a charging process in which each battery is charged at a constant current of 1.5A until a voltage of 4.2V is reached. Then it is charged in constant voltage mode until the charging current drops to 20mA. Secondly, there is a discharge process in which each battery is discharged at a constant current of 2A until the voltage of each battery drops to 2.5V. Finally, there is an impedance process in which electrochemical impedance spectroscopy is used to scan from 0.1 Hz to 5 kHz. The experimental conditions for different batteries are shown in Table 1.

[0083] Table 1 Different experimental conditions for different batteries

[0084] The experimental environment used in the embodiments of the present invention can be described as follows: hardware environment: AMD Ryzen 7 5800HCPU@3.20 GHz 3.20 GHz, NVIDIA GeForce RTX 3060 laptop GPU, 16 GB RAM, 64-bit operating system. The CNN-BiLSTM model is implemented in Python 3.9 using Pytorch. The present invention divides NASA's PCoE dataset into a training set and a test set. The data of the first 50% cycle of each battery is used as a training set, and the data of the second 50% cycle is used as a test set.

[0085] Hyperparameters are important factors in the CNN-BiLSTM model. The choice of hyperparameters may affect the performance and prediction effect of the network. The optimizer used in this example is Adam, the loss function is MSE loss, the activation function is LeakyReLU, the filter size is 5, the batch size is 2, the number of iterations is 600, the learning rate is 0.001, and the hidden size is 2.

[0086] This embodiment uses three error evaluation indicators as performance evaluation indicators, including mean square error (MSE), root mean square error (RMSE) and goodness of fit (R²). The smaller the mean square error (MSE) and root mean square error (RMSE), the greater the goodness of fit, indicating better prediction performance.

[0087]

[0088]

[0089]

[0090] In the formula, is the sample size; It is The actual value of samples; It is The predicted value of the sample, that is, the result predicted by the CNN-BiLSTM model; is the average of the actual values, that is, all The arithmetic mean of the values, are the actual values ​​of all samples.

[0091] At present, many advanced networks have been applied to the field of lithium-ion battery SOH prediction, and have achieved good results. The present invention selects two comparison methods, CNN and BiLSTM, for comparison. At the same time, compared with the method without sliding window input, the sliding window input of the original data is also used as a comparison method.

[0092] Method 1: CNN. CNN is one of the most commonly used neural networks and has been widely used in prediction tasks. CNN can adaptively extract high-dimensional abstract features, and the parameters of the convolution kernel are shared, which reduces network parameters and reduces the risk of overfitting.

[0093] Method 2: BiLSTM. LSTM is one of the most commonly used RNNs and is widely used in time series data prediction. BiLSTM is an improvement on the LSTM network. BiLSTM fully considers context information, solves the problem that traditional LSTM is sensitive to input sequences, and effectively prevents gradient vanishing and gradient exploding.

[0094] Method 3: CNN-BILSTM spatiotemporal dual-stream model proposed in the present invention. This embodiment adopts the fusion prediction method of CNN and BiLSTM. In addition to the above three comparison methods, this embodiment also uses two different learning strategies to compare the results of sliding window input and non-sliding window input. Therefore, this embodiment uses two learning strategies and three different networks for comparison, with a total of six comparison methods.

[0095] Learning strategy: sliding window. For the prediction of time series data, the sliding window operation of data is a common data preprocessing method. Usually several historical data are used to predict the future data of the next moment, which can make full use of all data.

[0096] Table 2 Prediction performance of different methods for B0005 and B0006 batteries

[0097] Compared with other methods, the results obtained by the method of the present invention have the lowest MSE and RMSE and the largest goodness of fit. For the B0005 battery, the MSE, RMSE and goodness of fit of the method of the present invention reached 0.00020, 0.01108 and 0.96750 respectively. For the B0006 battery, the MSE, RMSE and goodness of fit of the method of the present invention reached 0.00107, 0.03271 and 0.88880 respectively. This means that under the data of these two batteries, the prediction accuracy of the method of the present invention is the best among all methods. Under the same hyperparameters, the CNN-BILSTM spatiotemporal dual-stream model proposed in the present invention achieves the best prediction performance, and the prediction accuracy of the CNN-BILSTM spatiotemporal dual-stream model is far superior to all other methods.

[0098] from Figure 5 It can be seen from the results of the displayed features that different features have different importance for the prediction results of different periods. For example, features 1 and 2 are not important for the SOH prediction values ​​in the first 30 periods and the last 50 periods, but play an important role in predicting periods 30 to 120. Feature 3 accounts for a large proportion in the first 30 periods of SOH prediction, but its weight generally decreases in the following periods. Features 4 and 5 are important in predicting the first 30 periods and the last 20 periods, but are not so important in the middle stage of the prediction life cycle. The importance of features 6 and 7 gradually increases.

[0099] The above-mentioned embodiment of the present invention provides a robust and interpretable CNN-BiLSTM-based lithium-ion battery SOH prediction method, which uses 1D Grad-CAM++ to visualize the region of interest, thereby improving the interpretation ability of the CNN-BILSTM spatiotemporal dual-stream model. The method has been verified on the NASA PCoE dataset. The experimental results show that the method of the present invention can effectively predict the SOH of lithium-ion batteries, and the visualization results can help better understand the CNN-BILSTM spatiotemporal dual-stream model.

[0100] Step S6: preprocess the voltage, current and temperature data of the lithium-ion battery to be tested throughout its life cycle to obtain a lithium-ion battery capacity degradation sequence, input the lithium-ion battery capacity degradation sequence into the trained CNN-BILSTM spatiotemporal dual-stream model, and obtain a SOH prediction result of the lithium-ion battery.

[0101] The energy system life prediction method based on interpretable deep learning provided by the aforementioned embodiment of the present invention can predict the SOH of lithium-ion batteries. The spatiotemporal information of lithium-ion battery degradation is captured by the CNN-BILSTM spatiotemporal dual-stream model to improve the model feature extraction capability, and the SOH prediction is performed through a fully connected layer with prediction node output. After the model is trained, the output features are processed using an improved gradient class activation mapping technology (1D Grad-CAM++), and the region of interest is visualized and explained by combining with prior knowledge. The model is verified on the NASA lithium-ion battery dataset. The results show that the proposed method has better prediction performance and model interpretability than other existing technologies.

[0102] The energy system life prediction method provided by the aforementioned embodiments of the present invention can be used to predict the life of lithium-ion batteries, and can also be used to predict the life of other parts of the energy system. When predicting the life of other parts of the energy system, the input data of the CNN-BiLSTM spatiotemporal dual-flow model is changed accordingly. In some embodiments of the present invention, the energy system life prediction method is used to predict the life of an insulated gate bipolar transistor (IGBT).

[0103] As the sales of electric vehicles achieve rapid growth, insulated gate bipolar transistors (IGBTs) are the most popular components as key components for power conversion in chargers, ensuring the reliability of the charging process. In order to ensure the safe and reliable operation of electric vehicle chargers and reduce maintenance costs through reasonable system maintenance, the accurate remaining useful life (RUL) prediction of insulated gate bipolar transistors has become a trend issue to meet the electrical system's demand for higher reliability and safety.

[0104] In some embodiments of the present invention, the CNN-BILSTM spatiotemporal dual-stream model used in the above embodiments for lithium-ion battery life prediction is applied to NASA's IGBT degradation dataset to evaluate its generalization ability and prediction effect on different datasets. The experimental results are shown in Figure 2. Figure 6 , the CNN-BILSTM spatiotemporal dual-stream model also shows excellent performance on the IGBT dataset, and its predicted values ​​and actual label values ​​maintain a high degree of consistency in most cycles. Especially in the early stage of prediction, the CNN-BILSTM spatiotemporal dual-stream model can accurately capture the rapid decline trend of the remaining service life, which further verifies the sensitivity and accuracy of the CNN-BILSTM spatiotemporal dual-stream model to rapidly changing data. In addition, although the CNN-BILSTM spatiotemporal dual-stream model has some fluctuations in the later stage of prediction, it can still track the changes of actual values ​​well as a whole, which shows that the CNN-BILSTM spatiotemporal dual-stream model has good stability and robustness when dealing with long-term prediction tasks. These results not only prove the effectiveness of the CNN-BILSTM spatiotemporal dual-stream model in processing complex spatiotemporal data, but also demonstrate its generalization ability in different application scenarios. The CNN-BILSTM spatiotemporal dual-stream model can adapt to different data characteristics and changing trends, which is crucial to improving the accuracy and reliability of predictions. During actual detection, the degradation data of voltage and current of the insulated gate bipolar transistor throughout its life cycle are input into the trained CNN-BILSTM spatiotemporal dual-flow model to obtain the life prediction result of the insulated gate bipolar transistor.

[0105] In some embodiments of the present invention, a system for predicting the life of an energy system based on interpretable deep learning is provided, which is used to implement a method for predicting the life of an energy system based on interpretable deep learning. The system includes the following units: A detection unit is used to use a CNN-BiLSTM spatiotemporal dual-stream model to predict the life of the energy system; A model unit is used for training to obtain a CNN-BiLSTM spatiotemporal dual-stream model, wherein the CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module, wherein the deep feature extraction module includes a CNN feature extraction module, a BiLSTM feature extraction module and a fully connected layer, wherein the CNN feature extraction module is used to extract spatial features of the energy system test data, the BiLSTM feature extraction module is used to extract temporal features of the energy system test data, and the fully connected layer is used to predict the spliced ​​and fused temporal features and spatial features, and output the prediction results of the energy system; The interpretability module is used to visualize the output of the CNN feature extraction module through the 1D Grad-CAM++ feature visualization technology, map the output of the CNN feature extraction module back to the dimension of the energy system data to be tested, and output a class activation map reflecting the importance of the feature.

[0106] When used to predict the life of lithium-ion batteries, the data to be measured in the energy system is the capacity degradation sequence of the lithium-ion batteries. When used to predict the life of insulated gate bipolar transistors, the data to be measured in the energy system is the degradation data of the voltage and current of the insulated gate bipolar transistors over their entire life cycle. Of course, the method and system provided by the present invention are not limited to predicting the life of lithium-ion batteries and insulated gate bipolar transistors, but can also predict the life of other energy systems or other components of energy systems.

[0107] In some embodiments of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the energy system life prediction method based on explainable deep learning when executing the computer program.

[0108] In some embodiments of the present invention, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the energy system life prediction method based on explainable deep learning is implemented.

[0109] In summary, the results of the present invention fully demonstrate the strong potential and broad application prospects of the CNN-BILSTM spatiotemporal dual-stream model in energy system life prediction tasks. Whether in lithium-ion batteries or other energy systems such as IGBTs, the CNN-BILSTM spatiotemporal dual-stream model can provide high-quality prediction results. In addition, the CNN-BILSTM spatiotemporal dual-stream model can input both physical features and raw data, indicating that the CNN-BILSTM spatiotemporal dual-stream model has a wide range of applications and is highly flexible and adaptable.

[0110] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An energy system life prediction method based on explainable deep learning, characterized by: Input the data to be tested of the energy system into the CNN-BiLSTM spatiotemporal dual-stream model to obtain the life prediction results of the energy system; The CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module. The deep feature extraction module includes a CNN feature extraction module, a BiLSTM feature extraction module and a fully connected layer. The CNN feature extraction module is used to extract the spatial features of the data to be tested, and the BiLSTM feature extraction module is used to extract the temporal features of the data to be tested. The fully connected layer is used to predict the spliced ​​and fused temporal features and spatial features, and output the life prediction results of the energy system. The interpretability module visualizes the output of the CNN feature extraction module through 1D Grad-CAM++ feature visualization technology, and maps the output of the CNN feature extraction module back to the dimension of the data to be tested.

2. The energy system life prediction method based on interpretable deep learning according to claim 1 is characterized in that: The energy system includes a lithium-ion battery or an insulated gate bipolar transistor.

3. The energy system life prediction method based on interpretable deep learning according to claim 2 is characterized in that: When predicting a lithium-ion battery, the data to be measured in the energy system is a lithium-ion battery capacity degradation sequence that is sensitive to battery capacity and is obtained based on voltage, current and temperature data over the entire life cycle of the lithium-ion battery.

4. The energy system life prediction method based on interpretable deep learning according to claim 2 is characterized in that: When predicting an insulated gate bipolar transistor, the data to be measured of the energy system is the degradation data of the voltage and current of the insulated gate bipolar transistor over its entire life cycle.

5. The energy system life prediction method based on interpretable deep learning according to claim 1 is characterized in that: The CNN feature extraction module includes a convolution layer, an activation function, and an average pooling layer. The formula of the CNN feature extraction module is as follows: in, Represents the abstract spatial features extracted by the CNN feature extraction module, Indicates the data to be tested. represents the activation function, represents the weight, Indicates deviation.

6. The energy system life prediction method based on interpretable deep learning according to claim 1 is characterized in that: The BiLSTM feature extraction module includes a BiLSTM layer and a Flatten layer. The BiLSTM layer includes multiple LSTM layers. Each LSTM layer includes a forward LSTM and a reverse LSTM. The BiLSTM layer outputs a time feature extracted from the data to be tested. The time feature is a three-dimensional tensor. The Flatten layer is used to convert the three-dimensional tensor into a two-dimensional tensor.

7. The energy system life prediction method based on interpretable deep learning according to any one of claims 1 to 6, characterized in that: The formula of 1D Grad-CAM++ feature visualization technology is as follows: in, is the weight of the CNN feature extraction module, indicating the The prediction value of the final life span The weight of It is The weight of each feature; It is the output of the CNN feature extraction module, which refers to the input feature In the Channels and The value of the position; is the Grad-CAM++ heat map; It refers to The value of a feature.

8. The energy system life prediction system based on explainable deep learning is characterized by: The energy system life prediction method for implementing the explainable deep learning described in any one of claims 1 to 7, wherein the prediction system comprises the following units: A detection unit is used to use a CNN-BiLSTM spatiotemporal dual-stream model to predict the life of the energy system; A model unit, used for training a CNN-BiLSTM spatiotemporal dual-stream model, wherein the CNN-BiLSTM spatiotemporal dual-stream model includes a deep feature extraction module and an interpretability module; The deep feature extraction module includes CNN feature extraction module, BiLSTM feature extraction module and full connection layer. CNN feature extraction module is used to extract the spatial features of the energy system test data; BiLSTM The LSTM feature extraction module is used to extract the time features of the energy system data to be tested; the fully connected layer is used to predict the spliced ​​and fused time features and spatial features, and output the prediction results of the energy system; The interpretability module is used to visualize the output of the CNN feature extraction module through the 1D Grad-CAM++ feature visualization technology, and map the output of the CNN feature extraction module back to the dimension of the energy system data to be tested.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the energy system life prediction method based on explainable deep learning as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the energy system life prediction method based on explainable deep learning as described in any one of claims 1 to 7 is implemented.

Images