Multi-mode fusion switch cabinet temperature inversion method and device, and storage medium

Abstract

The invention relates to the field of power equipment monitoring, in particular to a multi-mode fusion switch cabinet temperature inversion method and device and a storage medium, and the method comprises the steps: obtaining the multi-mode working parameters of a switch cabinet; constructing feature vectors by using the multi-mode working parameters of the switch cabinet; constructing an LSTM model, and performing pre-training to obtain a trained model; and inputting the feature vector into a trained model to obtain a switch cabinet temperature inversion result. The beneficial effects of the invention are that the monitoring cost of the temperature of the switch cabinet is reduced, and the monitoring efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the field of power equipment monitoring, and in particular to a multi-modal fusion switch cabinet temperature inversion method, equipment and storage medium. Background Art

[0002] As the capacity and complexity of power systems continue to increase, the reliability and safety requirements for power equipment are also getting higher and higher. As one of the key equipment in the power system, the health status of the switchgear directly affects the stability of the entire power system. Traditional temperature monitoring methods often rely on direct contact sensors, which are not only costly but also inconvenient to install, especially on equipment that has been put into operation. In addition, contact sensors make it difficult to fully monitor the temperature distribution inside the switchgear and cannot detect potential overheating problems in a timely manner. Summary of the invention

[0003] The purpose of the present invention is to propose a multi-modal fusion switch cabinet temperature inversion method, device and storage medium to solve the technical problems of incomplete monitoring and high cost caused by traditional switch cabinet temperature monitoring relying on contact sensors.

[0004] The present invention provides a multi-mode fusion switch cabinet temperature inversion method, comprising the following steps:

[0005] S1. Obtain multi-mode operating parameters of the switch cabinet;

[0006] S2, constructing feature vectors using multi-modal operating parameters of the switch cabinet;

[0007] S3. Build an LSTM model and pre-train it to obtain a trained model;

[0008] S4. Input the feature vector into the trained model to obtain the switch cabinet temperature inversion result.

[0009] A storage medium stores instructions and data for implementing a multi-modal fusion switch cabinet temperature inversion method.

[0010] A multi-modal fusion switch cabinet temperature inversion device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a multi-modal fusion switch cabinet temperature inversion method.

[0011] The beneficial effects provided by the present invention are:

[0012] Multimodal data fusion: The present invention improves the accuracy and robustness of temperature inversion by fusing data of multiple modes such as the surface temperature of the switch cabinet, operating condition information (such as current, voltage), working mode information, and other environmental parameters. The use of multimodal data fusion technology can more comprehensively reflect the actual operating status of the switch cabinet, thereby more accurately predicting the internal temperature distribution. Application of deep learning model: The present invention uses the LSTM model to process time series data. The model can capture long-term dependencies in the sequence and is particularly suitable for processing time series problems such as temperature inversion. By using the LSTM model, the changing trend of the internal temperature of the switch cabinet can be better simulated, thereby improving the prediction accuracy.

[0013] Mathematical expression of working mode: This invention accurately describes the different working modes of the switch cabinet by defining discrete variables and logical expressions. These mathematical formulas and logical expressions are used as one of the input parameters to help the algorithm more accurately predict the temperature distribution of the switch cabinet in different working modes. The mathematical expression of the working mode enables the algorithm to better understand the behavior of the switch cabinet in different states, thereby improving the accuracy of temperature prediction.

[0014] Reconstruction of temperature field distribution: The present invention obtains the distributed temperature of the main internal components by placing several more temperature sensors during the test phase, and reconstructs the temperature field distribution by using the temperature proportional relationship between these sensors. In this way, even if only one sensor is used in actual application, the temperature distribution inside the switch cabinet can be reconstructed, thereby reducing the monitoring cost and improving the monitoring efficiency.

[0015] These innovations together constitute an efficient and accurate switchgear internal temperature inversion algorithm, which can provide strong support for the operation and maintenance of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flow chart of the method of the present invention;

[0017] Figure 2 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0019] First, some basic explanations of the relevant concepts are given.

[0020] Switchgear is a kind of equipment used in power system, which is mainly used to control, protect and isolate power lines.

[0021] Please refer to Figure 1 , Figure 1: is a flow chart of the method of the present invention; the present invention provides a multi-modal fusion switch cabinet temperature inversion method, the method specifically includes the following steps:

[0022] S1. Obtain multi-mode operating parameters of the switch cabinet;

[0023] It should be noted that in the present invention, various parameters of the switch cabinet involved in the temperature inversion calculation need to be normalized for data processing and feature extraction for the following reasons:

[0024] 1) Scale consistency: Different physical quantities (such as temperature, voltage, and current) have different dimensions and value ranges. If normalization is not performed, during model training, features with larger value ranges may dominate the model's learning process, while features with smaller values ​​may be ignored.

[0025] 2) Accelerate convergence: Normalization can speed up the convergence of machine learning models, especially when using gradient descent for optimization, the normalized features help reduce the number of iterations.

[0026] 3) Avoid gradient explosion / vanishing: For deep learning models, normalization can avoid the problem of gradient explosion or gradient vanishing, which helps the model learn more stably.

[0027] Specifically, the multi-modal operating parameters of the switch cabinet in step S1 include: surface temperature parameters, operating condition parameters, operating modes and switch action parameters.

[0028] The surface temperature parameters include: upper surface temperature parameters, lower surface temperature parameters, average temperature and temperature change rate.

[0029] Regarding the surface temperature parameters, the specific process of obtaining the parameters in the present invention is as follows:

[0030] A temperature sensor is installed on the outer surface of the upper and lower doors on the back of the switch cabinet by magnetic attraction. The temperature data of the upper and lower surfaces of the switch cabinet can be expressed as a time series, T upper (n) and T lower (n) represents the temperature of the upper surface and the lower surface respectively. An ambient temperature sensor is installed on the pole one meter away from the switch cabinet. T en (n), n is the time step, and its value range is n=1,2,...N.

[0031] Normalization:

[0032]

[0033] Where min and max represent the minimum and maximum values ​​in the entire time series, respectively.

[0034] Average temperature:

[0035]

[0036] Temperature change rate:

[0037]

[0038] Regarding the operating parameters, the process of obtaining the parameters in the present invention is as follows:

[0039] The load current and voltage are collected by current transformers and voltage transformers, and the time series data of the load current and voltage are obtained, where I(n) and U(n) represent the load current and voltage respectively. Similarly, n is the time step, and the value range is n=1, 2, ...N.

[0040] a. Voltage normalization:

[0041]

[0042] b. Current normalization:

[0043]

[0044] Where min and max represent the minimum and maximum values ​​in the entire time series, respectively.

[0045] Voltage mean:

[0046] Voltage change rate:

[0047] Current mean:

[0048] Current change rate:

[0049] Regarding the working mode, the acquisition process of the present invention is as follows:

[0050] The switchgear can adopt a variety of working modes according to different operating requirements and power system configuration. The following are several common switchgear working modes:

[0051] (1) Operation mode: In this mode, the circuit breaker in the switch cabinet is in the closed state and power can pass normally. The grounding knife switch (if any) should be in the open state to prevent accidental grounding.

[0052] (2) Hot standby mode: The circuit breaker is in the open state, but the isolating switch (if any) is in the closed state. This mode allows quick power-on because the isolating switch does not need to be operated. The grounding switch is in the open state.

[0053] (3) Cold standby mode: The circuit breaker and the isolating switch are both in the open state. Usually used before planned maintenance work or when a circuit is not to be used. The grounding switch can be closed for safe maintenance operations.

[0054] (4) Maintenance mode: All switches (including circuit breakers and isolating switches) are in the open state. The grounding switch is closed to ensure that maintenance personnel will not touch live parts when performing work. Other safety measures may also be taken, such as hanging warning signs.

[0055] (5) Isolation mode: The circuit breaker is in the open state and the isolation switch is also open to completely isolate the circuit. The grounding switch can be closed to ensure that there is no residual voltage. This mode is used when a circuit needs to be completely separated from the power grid.

[0056] (6) Test mode: In this mode, the circuit breaker and the isolating switch are disconnected and the grounding switch is closed. It is used to test the function and performance of the equipment to ensure that all components can work properly.

[0057] In order to express the different working modes of the switch cabinet with mathematical formulas and use them for the multi-modal fusion switch cabinet temperature inversion algorithm, the present invention defines a set of discrete variables to represent each mode and uses logical expressions to describe the switch state. The following is a mathematical formula for each working mode:

[0058] a. Operation mode:

[0059] Circuit breaker closed: CB = 1;

[0060] The grounding switch is disconnected: ES=0;

[0061] Mode flag: M_op=1;

[0062]

[0063] b. Hot standby mode:

[0064] Circuit breaker disconnected: CB = 0;

[0065] Isolation switch closed: IS=1;

[0066] The grounding switch is disconnected: ES=0;

[0067] Mode flag: M_hot_standby = 1;

[0068]

[0069] c. Cold standby mode:

[0070] Circuit breaker disconnected: CB = 0;

[0071] Isolation switch disconnected: IS=0;

[0072] The grounding switch can be closed: ES = 1 or ES = 0;

[0073] Mode flag: M_cold_standby=1;

[0074]

[0075] d. Maintenance mode:

[0076] All switches are open: CB = 0 and IS = 0;

[0077] The grounding switch is closed: ES=1;

[0078] Mode flag: M_maintenance = 1;

[0079]

[0080] e. Isolation mode:

[0081] Circuit breaker disconnected: CB = 0;

[0082] Isolation switch disconnected: IS=0;

[0083] The grounding switch can be closed: ES = 1;

[0084] Mode identification: M_isolation = 1;

[0085]

[0086] f. Test mode:

[0087] Circuit breaker disconnected: CB = 0;

[0088] Isolation switch disconnected: IS=0;

[0089] The grounding switch is closed: ES=1;

[0090] Mode identification: M_test=1;

[0091]

[0092] Note that in the maintenance mode, isolation mode and test mode, the grounding switch is closed, but their purposes and contexts are different. In practical applications, these modes are usually mutually exclusive, that is, only one mode can be active at the same time. Therefore, in the actual algorithm implementation, it can be ensured that only one of these mode identification variables (M_op, M_hot_standby, M_cold_standby, M_maintenance, M_isolation, M_test) is 1, and the rest are 0.

[0093] These mathematical formulas and logical expressions can be used in the multi-modal fusion switch cabinet temperature inversion algorithm as one of the input parameters to help the algorithm more accurately predict the temperature distribution of the switch cabinet in different operating modes.

[0094] Regarding the switch action parameters, the acquisition process of the present invention is as follows:

[0095] As an important part of switchgear operation, the switch action can be combined with the previously described working mode to provide more comprehensive status information. The following is a description of the switch action and links it with the switch working mode. The switch action sequence S(n) can be defined as:

[0096] S(n)=(t n ,type n )for n=1,2,....N,type n ∈{close, open} (21)

[0097] where t n It is a timestamp, indicating the time when the switch action occurs. n It is the switch type, indicating whether it is a closing operation or an opening operation.

[0098] Since the switching actions occur over time and may span multiple operating modes, a more practical approach is to process the switching action sequences as time series data in parallel with the operating modes.

[0099] In practical applications, the operating mode of the switchgear can be inferred based on the sequence of switching actions. For example, if a series of closing and opening actions are detected and these actions match a specific time interval and sequence pattern, it can be inferred that the switchgear may be in a specific operating mode (such as hot standby mode or test mode).

[0100] Therefore, while the switching actions themselves do not directly define the operating mode, they provide important information about the operating state of the switchgear and can be used to infer and verify the operating mode.

[0101] In the switchgear temperature inversion algorithm based on multi-modal fusion, the switch action sequence can be used as an additional input parameter to help the algorithm more accurately predict and explain the temperature behavior of the switchgear.

[0102] S2, constructing feature vectors using multi-modal operating parameters of the switch cabinet;

[0103] It should be noted that, in combination with the above parameters, the feature vector constructed by the present invention is as follows:

[0104]

[0105] S3. Build an LSTM model and pre-train it to obtain a trained model;

[0106] It should be noted that the LSTM (Long Short-Term Memory) model is a special recurrent neural network (RNN) designed to solve the gradient vanishing and gradient exploding problems encountered by traditional RNNs when processing long time series data. LSTM introduces a "gate" mechanism to control the flow of information, enabling the model to learn long-term dependencies. LSTM is a powerful recurrent neural network variant that is particularly suitable for processing time series data. By introducing a gating mechanism, LSTM is able to effectively handle long-term dependencies and is more stable during training. These features make it one of the ideal choices for dealing with multimodal temperature inversion problems, especially when it is necessary to capture patterns and dependencies in time series data.

[0107] By using the LSTM model and multimodal data fusion, we can build a model that can predict the temperature distribution inside the switchgear. During the initial training phase, the prediction accuracy of the model can be further improved by installing internal temperature sensors and using this data as a reference standard. This approach can not only help monitor the temperature conditions inside the switchgear, but also detect potential safety hazards in a timely manner, thereby improving the reliability and safety of the power system.

[0108] Step S3 in the present invention is specifically as follows:

[0109] S31. Build an LSTM model, whose objective function is as follows:

[0110]

[0111] Where T int_ref (i) is the temperature value measured by the temperature sensor inside the switch cabinet during the training phase (only used during the training phase), T int (i) is the internal temperature value predicted by the model; N is the number of samples;

[0112] It should be noted that the present invention uses the LSTM model for training, the input is X(n), and the output is the target temperature T int (n), the objective function uses mean square error (MSE).

[0113] S32. Use the Adam optimizer to solve the objective function and minimize it to obtain the trained model parameters.

[0114] It should be noted that common optimizers include Adam, RMSprop or SGD. Here we introduce the solution process using the Adam optimizer. Adam (Adaptive Moment Estimation) is a widely used deep learning optimization algorithm.

[0115] Specifically, step S32 is as follows:

[0116] S321, initialize model parameters ω0, initialize first-order moment estimate m0=0 and second-order moment estimate ν0=0; set initial learning rate α; set first-order moment attenuation factor β1; set second-order moment attenuation factor β2;

[0117] S322. For each time step n, calculate the loss function with respect to the parameter ω n The gradient g n ;

[0118] It should be noted that in actual operations, these calculations are usually automatically completed by deep learning frameworks (such as TensorFlow or PyTorch), and users only need to define the model structure and loss function.

[0119] S323, update first-order moment estimate: m n =β1m n-1 +(1-β1)g n ; This is a weighted average, where g n The weight of the gradient gradually increases, and the weight of the old gradient gradually decreases;

[0120] S324, update the second-order moment estimate: ν n =β2ν n-1 +(1-β2)g n 2 ; This is also a weighted average used to estimate the variance of the gradient;

[0121] S325, deviation correction:

[0122] It should be noted that since the first-order moment and second-order moment estimates at the initial moment may be biased towards zero, bias correction is required.

[0123] S326. Update model parameters: Used to estimate the standard deviation of the gradient, ε is a decimal used to avoid division by zero;

[0124] S327. Use the parameter adjustment method to set the stop parameter update condition to obtain the final trained model parameters.

[0125] It should be noted that there are several methods for adjusting parameters:

[0126] Hyperparameter tuning: In addition to the weights and biases that are automatically adjusted by back-propagation, some hyperparameters need to be manually tuned, such as learning rate, batch size, number of hidden layer units, etc. The best combination can be found using methods such as grid search or random search.

[0127] Early stopping method: If the performance on the validation set no longer improves, the training is terminated early to prevent overfitting.

[0128] Regularization: You can use regularization such as L1 or L2 to reduce the risk of overfitting.

[0129] The following uses the Early Stopping method as an example to introduce parameter adjustment. Early Stopping is a commonly used technique to prevent overfitting. It determines when to stop training by monitoring the performance on the validation set.

[0130] How early stopping works:

[0131] Define a validation set: A portion of the training data is used as a validation set to monitor the performance of the model on unseen data.

[0132] Set monitoring metrics: Choose a performance metric, such as a loss function on a validation set (such as mean squared error) or other evaluation metrics (such as accuracy).

[0133] Set a threshold and waiting period: Set a threshold. When the performance change on the validation set is less than this threshold, the performance is considered no longer significantly improved. Set a waiting period (patience), that is, stop training after the performance has not improved for a certain number of consecutive periods.

[0134] Training process: After each cycle, check the performance on the validation set. If the performance has not improved, record the count of waiting cycles. If the performance has improved, reset the count of waiting cycles. When the count of waiting cycles reaches the preset threshold, stop training.

[0135] Example of early stopping implementation:

[0136] Below is a simple code example of implementing the early stopping method using TensorFlow and Keras. Suppose you are building a simple LSTM model to predict the temperature distribution inside a switch cabinet.

[0137] Create a model: A simple LSTM model is defined, which contains two LSTM layers and an output layer.

[0138] Compile the model: Use mean squared error as the loss function and Adam as the optimizer.

[0139] Define the early stopping callback: monitor = 'val_loss' indicates that the loss on the validation set will be monitored. patience = 10 indicates that if the loss on the validation set does not improve significantly for 10 consecutive epochs, the training will be stopped. restore_best_weights = True indicates that at the end of training, the model's weights will be restored to the state of the epoch with the best performance on the validation set.

[0140] Training model: Use training data set X_train and y_train for training. Use validation data set X_val and y_val for validation.

[0141] Use Early Stopping to prevent overfitting.

[0142] By using the early stopping method, the model can be effectively prevented from overfitting during the training process. In practical applications, the value of patience can be adjusted according to the specific situation to balance the model performance and training time. In addition, other techniques (such as regularization) can be combined to further improve the generalization ability of the model.

[0143] S4. Input the feature vector into the trained model to obtain the switch cabinet temperature inversion result.

[0144] At this point, the final adjusted model is applied to actual scenarios, such as real-time monitoring of the temperature distribution inside the switch cabinet.

[0145] Finally, the method of the present invention is summarized as follows:

[0146] 1) Data collection and preprocessing: Collect switchgear surface temperature, current, voltage data and working mode information. Perform data normalization and feature extraction.

[0147] 2) Feature extraction and model building: Extract useful features such as average temperature, temperature change rate, current and voltage change rate, etc. Build an LSTM model and use the Adam optimizer for training.

[0148] 3) Parameter adjustment and model training: Hyperparameter tuning, early stopping and regularization techniques are used to prevent overfitting and improve the generalization ability of the model. For each epoch, forward propagation is performed on the training set, the loss is calculated, backpropagation and gradient descent are performed using the Adam optimizer to update the model parameters, and the model performance is evaluated on the validation set.

[0149] 4) Error minimization: Through parameter adjustment during training, the model gradually learns how to minimize the difference between the predicted temperature and the actual temperature. At the end of each epoch, the model updates the parameters to minimize the loss function on the validation set. Once the performance on the validation set no longer improves, the training is terminated early through early stopping to prevent overfitting.

[0150] 5) Application model: Apply the final adjusted model to the actual scenario and monitor the temperature distribution inside the switch cabinet in real time.

[0151] Applying the model of the present invention to obtain distributed temperature In order to obtain the distributed temperature of the main components inside the switch cabinet, the following steps are taken:

[0152] 1) Temperature sensor arrangement during the test phase: During the test phase, multiple temperature sensors are placed around the main components inside the switch cabinet to collect temperature data at different locations. These sensors can be distributed at key locations inside the switch cabinet, such as circuit breakers, busbars, etc. The temperature data of these sensors are collected, and the temperature ratio relationship between each sensor is recorded.

[0153] 2) Model training and temperature inversion: Use the trained model to predict the temperature of a certain point inside the switch cabinet. Use the temperature ratio collected during the test phase to estimate the temperature at other locations using the predicted temperature values. In this way, the temperature distribution inside the switch cabinet can be reconstructed even if only one sensor is used in actual applications.

[0154] 3) Reconstruction of temperature field distribution map: Based on the predicted temperature value and the proportional relationship in the test phase, the temperature field distribution map inside the switch cabinet can be drawn. This method not only reduces the number of sensors required for real-time monitoring, but also provides visual information on the temperature distribution inside the switch cabinet.

[0155] 4) Real-time monitoring and application: In actual applications, only one sensor temperature needs to be monitored, and the proportional relationship established during the test phase is used to estimate the temperature at other locations. In this way, the temperature distribution inside the switch cabinet can be monitored in real time, and potential safety hazards can be discovered in a timely manner.

[0156] Please refer to Figure 2 , Figure 2It is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a multi-modal fusion switch cabinet temperature inversion device 401, a processor 402 and a storage medium 403.

[0157] A multi-modal fusion switch cabinet temperature inversion device 401: The multi-modal fusion switch cabinet temperature inversion device 401 implements the multi-modal fusion switch cabinet temperature inversion method.

[0158] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the switch cabinet temperature inversion method of multi-modal fusion.

[0159] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the switch cabinet temperature inversion method of multi-modal fusion.

[0160] In summary, the beneficial effects of the present invention are:

[0161] Multimodal data fusion: The present invention improves the accuracy and robustness of temperature inversion by fusing data of multiple modes such as the surface temperature of the switch cabinet, operating condition information (such as current, voltage), working mode information, and other environmental parameters. The use of multimodal data fusion technology can more comprehensively reflect the actual operating status of the switch cabinet, thereby more accurately predicting the internal temperature distribution. Application of deep learning model: The present invention uses the LSTM model to process time series data. The model can capture long-term dependencies in the sequence and is particularly suitable for processing time series problems such as temperature inversion. By using the LSTM model, the changing trend of the internal temperature of the switch cabinet can be better simulated, thereby improving the prediction accuracy.

[0162] Mathematical expression of working mode: This invention accurately describes the different working modes of the switch cabinet by defining discrete variables and logical expressions. These mathematical formulas and logical expressions are used as one of the input parameters to help the algorithm more accurately predict the temperature distribution of the switch cabinet in different working modes. The mathematical expression of the working mode enables the algorithm to better understand the behavior of the switch cabinet in different states, thereby improving the accuracy of temperature prediction.

[0163] Reconstruction of temperature field distribution: The present invention obtains the distributed temperature of the main internal components by placing several more temperature sensors during the test phase, and reconstructs the temperature field distribution by using the temperature proportional relationship between these sensors. In this way, even if only one sensor is used in actual application, the temperature distribution inside the switch cabinet can be reconstructed, thereby reducing the monitoring cost and improving the monitoring efficiency.

[0164] These innovations together constitute an efficient and accurate switchgear internal temperature inversion algorithm, which can provide strong support for the operation and maintenance of the power system.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-modal fusion switch cabinet temperature inversion method, characterized by: The following steps are involved: S1. Obtain multi-mode operating parameters of the switch cabinet; S2, constructing feature vectors using multi-modal operating parameters of the switch cabinet; S3. Build an LSTM model and pre-train it to obtain a trained model; S4. Input the feature vector into the trained model to obtain the switch cabinet temperature inversion result.

2. The multi-modal fusion switch cabinet temperature inversion method according to claim 1, characterized in that: The multi-modal operating parameters of the switch cabinet in step S1 include: surface temperature parameters, operating condition parameters, operating mode and switch action parameters.

3. The multi-modal fusion switch cabinet temperature inversion method according to claim 2, characterized in that: The surface temperature parameters include: upper surface temperature parameters, lower surface temperature parameters, average temperature and temperature change rate.

4. The multi-modal fusion switch cabinet temperature inversion method according to claim 2, characterized in that: The operating parameters include: voltage parameters and current parameters; wherein the voltage parameters include voltage mean value and voltage change rate; and the current parameters include current mean value and current change rate.

5. The multi-modal fusion switch cabinet temperature inversion method according to claim 2, characterized in that: The working modes include: operation mode, hot standby mode, cold standby mode, maintenance mode, isolation mode and test mode.

6. The multi-modal fusion switch cabinet temperature inversion method according to claim 2, characterized in that: Step S3 is as follows: S31. Build an LSTM model, whose objective function is as follows: Where T int_ref (i) is the temperature value measured by the temperature sensor inside the switch cabinet during the training phase, T int (i) is the internal temperature value predicted by the model; N is the number of samples; S32. Use the Adam optimizer to solve the objective function and minimize it to obtain the trained model parameters.

7. A multi-modal fusion switch cabinet temperature inversion method according to claim 6, characterized in that: Step S32 is specifically as follows: S321, initialize model parameters ω0, initialize first-order moment estimate m0=0 and second-order moment estimate ν0=0; set initial learning rate α; set first-order moment attenuation factor β1; set second-order moment attenuation factor β2; S322. For each time step n, calculate the loss function with respect to the parameter ω n The gradient g n ; S323, update first-order moment estimate: m n =β1m n-1 +(1-β1)g n ; S324, update the second-order moment estimate: ν n =β2ν n-1 +(1-β2)g n 2 ; S325, deviation correction: S326. Update model parameters: Used to estimate the standard deviation of the gradient, ε is a decimal used to avoid division by zero; S327. Use the parameter adjustment method to set the stop parameter update condition to obtain the final trained model parameters.

8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a multi-modal fusion switch cabinet temperature inversion method as described in any one of claims 1 to 7.

9. A multi-modal fusion switch cabinet temperature inversion device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a multi-modal fusion switch cabinet temperature inversion method as described in any one of claims 1 to 7.

Images