In order to facilitate the extraction of high-dimensional features of data, the Inception module is used to build a CNN model.

Inception module is the basic component of GoogLeNet. This module decomposes the larger convolution layer in CNN into smaller convolution layer and pooling layer, which can improve the performance in the following two aspects: First, network learning parameters are reduced through the decomposition of the convolution layer, and the model learning efficiency is improved. Second, the decomposition of the convolution layer enables the model to have the ability of multi-channel information processing, which improves the breadth of the network. Since the model outputs the summary of multi-channel processing results, the information representation ability is enhanced. The specific operation of convolution kernel decomposition is as follows:

(1) Convolution kernel decomposition. 5 × 5 convolution is decomposed into two serial 3 × 3 convolution with the same receptive field, while the latter has more non-linear operations. The number ratio of the two parameters is 3 2 + 3 2 / 5 2 = 0.72 , that is, the number of parameters is reduced by 28%.

The form of convolution decomposition can be seen in the Inception module structure shown in Figure 1B. As can be seen from Figure 1B, Inception module expands both the breadth and depth of the network. That is, pooling operation and multi-scale convolution processing are adopted to process data. Such multi-channel data processing method increases the network breadth. The decomposition of convolution kernel increases the depth of the network, and also increases the number of activation functions, thus strengthening the characterization ability of the network.

Power switch classification problem requirement model to predict all the state of the switch in the power distribution network, to reduce the pressure of the classification of the Softmax classifier, this paper takes the IEEE 33 buses system as an example, based on the loop structure of distribution network structures, multiple single loop CNN submodel, combining classifier output of each sub models predict each state of the switch. The structure of CNN sub-model is shown in Figure 1C.

As can be seen from Figure 1C, the model first preprocesses the input distribution network parameter matrix rows with convolution and pooling, and then imports the feature extraction part composed of Inception module. The data features are transmitted to the full connection layer after secondary sampling and flattening for mapping be-tween data and labels. Finally, Softmax function is used to calculate and output the state prediction results of each switch in the corresponding loop.

Loop switch state prediction is a multi-classification problem, and its loss optimization process is as follows:

(1) The Softmax classifier outputs the classification index. The Softmax function is defined as the ratio of the exponent of a single element to the exponent sum of all elements:

Adam algorithm is the combination of momentum acceleration decline and root mean square back propagation. By calculating the first and second moment estimates of the gradient, independent adaptive learning rates are designed for different parameters. The algorithm can correct the problems of learning rate disappearance and slow con-vergence in the optimization process, and its iterations are as follows: g t = ∇ θ L t θ t − 1 m ^ t = β 1 m t − 1 + 1 − β 1 g t 1 − β 1 t v ^ t = β 2 v t − 1 + 1 − β 1 g t 2 1 − β 2 t θ t = θ t − 1 − α m ^ t v ^ t + ε t = t + 1 (3)

In the reconfiguration calculation of distribution network, the actions of switches between loops will affect each other. Therefore, the interaction between sub-models should also be considered in the training of CNN model for unified optimization. The structure of the integration model is shown in Figure 1D.

It can be seen from Figure 1D that the number of branches of the integration model is the same as the loop number L of the distribution network. During the data input phase of the model, a convolutional layer with Batch Normalization BN is set up for initial data processing, which is then fed into each circuit sub-model.

The Concatenate layer is used to combine the output of the model, and the final output of the model is the splicing of the output of sub-model classifiers in the first dimension: O con = O 1 , O 2 , . . . , O l (4)

The model output needs to be combined with the corresponding data labels to calculate the loss of each branch, and take the maximum value to participate in the weight iteration of the model. The loss is calculated as: L c = max L 1 , L 2 , . . . , L l (5)

For the loss L c calculated by the model, the iteration format of Adam algorithm’s parameters is: g t c = ∇ θ L t c θ t − 1 c m ^ l , t = β 1 m l , t − 1 + 1 − β 1 g t c 1 − β 1 t v ^ l , t = β 2 v l , t − 1 + 1 − β 1 g t c 2 1 − β 2 t θ l , t = θ l , t − α m ^ l , t v ^ l , t + ε t = t + 1 (6)

The dataset contains 19 DG scenarios as shown in Table 3, each of which is divided into 512 load modes and a total of 9,728 system state sections. In the training process, the proportion of training set to total data is 80%, test set to 15%, verification set to 5%. To verify the performance advantage of the CNN model based on Inception module in this paper, the training effect is compared with that of the basic CNN model. The accuracy of the two models on the training set and test set is shown in Figure 5.

It can be seen from Figure 5 that the Inception CNN model has better feature extraction capability than the basic convolutional model. The training effect of the former on the training set and test set and the overall accuracy increase rate of the former are better than the latter. After 250 rounds of training, the accuracy of the model in the training set and test set is 99.31% and 99.28% respectively. From the curve trend, in the first 50 rounds, the precision of the model training set is ahead of that of the test set. During the 50 to 250 rounds of training, the accuracy curves of the two are similar, and tend to be stable after 200 rounds. In the early stages of training (rounds 0–25), due to the differences among the branches of the Inception model, which enable the optimization of the test accuracy is delayed. Thereby, resulting in the classification accuracy of the Inception test set being lower than that of the basic Conv model test set. Due to the focus of the hybrid training method on the high-loss branch of the model, the final training effect of the model is guaranteed.

According to Figure 6, the main operation switches of system optimization recon-figuration are located at B7, B9, B14, B23, and B26. In addition, the classification error of the model is mainly reflected in the confusion of switch states between branches B9-B10-B14 and B22-B23.

In addition, the performance index of DL model is the quantification of model performance based on confusion matrix, including accuracy, recall rate and F1 score. Accuracy rate represents the real tag proportion of model prediction results. Recall rate represents the prediction accuracy of the model for real labels. The F1 score is the harmonic average of the first two, and the F1 score of the model is positively correlated with its performance. The classification index of action branch based on confusion matrix is shown in Table 4.

It can be seen from Table 4 that the switch classification prediction effect of the model for large sample branches and most small sample branches is good, and each performance score is above 99. However, the classification performance of small sample branches B10 and B22 decreased, and their F1 scores were 97.15 and 90.91, respectively. It shows that insufficient label sample size will lead to the decrease of model generalization ability. The reason is on the one hand, the data feature coverage of small sample is not complete. On the other hand, it is the coverage of large sample features to small sample features in data sets.

In order to verify the timeliness of THE CNN model, 2,000 groups of data were randomly selected from the data set as the prediction target, and imported into the CNN model completed by training for prediction and timing. The results of program running time are shown in Table 5. In Table 5, a and b are configured to simulate offline operation, and the model loading time is included in the total time consumed, which includes the model loading time and the program operation time. Configuration c is set to simulate online operation, and the total time consumed is only the operation time of the program.

By comparing the operation of each configuration in Table 5, it can be seen that the model loading time accounts for more than 65% of the total offline operation. Excluding model loading, the calculation time of all three is less than 2 ms, showing no significant difference. By comparing configuration A and B, it can be seen that when the amount of calculation data is increased, the proportion of model loading time decreases from 93.43% to 65.35%, and the influence on the total time is reduced.

The standard IEEE33-bus distribution network is taken as the test system. The proposed method is compared with QPSO, improved cuckoo algorithm improved cuckoo search algorithm (ICSA), random forest (RF) and long short-term memory (LSTM) to calculate the average of 50 metaheuristic algorithms. The solution results are shown in Table 6.

As shown in Table 6, compared with QPSO, ICSA, RF, and LSTM methods, the online and offline operation time of the proposed method is significantly reduced and controlled to within 1s in terms of solution efficiency. Therefore, the proposed method can realize the rapid reconfiguration of distribution network, because the CNN model does not need to perform iterative optimization in the solution space. The reconfiguration strategy can be directly obtained according to the mapping relationship between the load pattern established in the training process and the reconfiguration strategy. In terms of network loss, the loss reduction ability of all the five algorithms is similar. Among them, the network loss of the ICSA algorithm is the same as that of the proposed method, which is 139.554 KW, while the network loss of the other three methods is slightly higher but it can be maintained below 139.590 KW. Therefore, compared with other algorithms, the fast reconfiguration method of distribution network based on CNN proposed in this paper improves the solution efficiency and reduces the system network loss at the same time.

In conclusion, there is no significant difference between online and offline data calculation speed of CNN model, while the influence of model loading time on the overall calculation speed can be diluted by increasing single input data amount in offline calculation. In addition, the precision of data-driven model needs a large amount of data support, while the meta-heuristic algorithm has little data requirements, which is more advantageous in the optimization and reconfiguration of new systems. Therefore, CNN model is more suitable for the scenarios of online real-time reconfiguration and mass reconfiguration of data operations.

In order to verify the effect of CNN model in loss reduction proposed in this paper, 190 groups of data in the validation set were extracted for prediction, and network loss and voltage level before and after reconfiguration were compared and predicted. Figure 7A shows the comparison of network loss before and after reconfiguration.

According to Figure 7A, compared with the topology before system reconfiguration, the topology optimized by CNN has lower network active power loss. According to the loss reduction percentage curve, in the test data, the average network loss of the system is reduced by 35.63% after network reconfiguration. The reason is that CNN’s optimized re-selection of network open-loop points balances the power flow of each branch and reduces the active network loss of the system.

Further, in order to analyze the influence of CNN model proposed in this paper on voltage before and after reconfiguration. Figures 8A,B respectively show the comparison diagram of bus minimum voltage and voltage range before and after system re-configuration. Figure 7B shows the effect of the reconfiguration on the system voltage level.

It can be seen from Figures 8A,B that the optimization of network bus voltage by CNN prediction topology is reflected in the increase of bus minimum voltage and the decrease of bus voltage range. According to Figure 7B, in the test data, the voltage range decreased by 35.47% on average, and the minimum voltage of system buses increased by 1.76% on average. The reason is that CNN prediction topology optimizes the power flow of the system, reduces the voltage loss of each branch, reduces the voltage difference between buses, improves the power supply quality of the system, and contributes to the low-carbon economic operation of the distribution network.

In order to further verify the effectiveness of the fast reconfiguration method of low-carbon distribution network based on CNN in large-scale distribution systems, a 185-bus distribution system is taken as a test system. The system has 185 buses, 184 branches and 75 switches, including 20 tie switches and 55 segment switches. The parameters of DGs are shown in Table 7.

The results of program running time are shown in Table 8. As can be seen, due to the complexity of the 185-bus test system, the CNN model requires additional parameters to describe the relationship between load patterns and reconfiguration strategies, resulting in model data loading times and mean times longer than the IEEE33 node system, ultimately increasing the solution time. However, even though the test system is more complex, the mean solution time of the proposed CNN model remains below 2.5 ms and rapid reconfiguration is achieved.

Running results of 185-bus system are shown in Table 9. It can been seen in Table 9, compared with the original network, CNN reduces the power loss by 22.86% while improving the lowest node voltage from 0.9253p.u. to 0.9508 p.u. As a result, the CNN method proposed in this paper also has good performance advantages in the 185-node system.