Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 12, No. 3, 2024 

 

125 

SF6	Gas	Humidity	Prediction	Model	Based	on	Deep	
Learning	
Xiaomin Chen1, *, Zhangyun Li2 

1Guangzhou Power Supply Bureau of Guangdong Power Grid Co., Ltd, Guangzhou 510000, China 
2Electric Power Science Research Institute of Southern Power Grid Ultra High Voltage Transmission Company, Guangzhou 510000, China 
* Corresponding author 

 

Abstract: SF6 gas is widely used in Gas Insulated Switchgear(GIS) as an insulating and arc extinguishing medium in electric 
power industry. However, SF6 gas humidity has a significant impact on the performance and reliability of GIS equipment. The 
accurate prediction of humidity level is one of the keys to ensure the long-term stable operation of GIS equipment. Traditional 
humidity prediction methods are often limited by model complexity and data processing ability, which is difficult to meet the 
actual demand. In this paper, deep learning, as a powerful data-driven approach, shows great potential in gas humidity prediction. 
An efficient SF6 gas humidity prediction model for GIS equipment is constructed based on deep learning algorithm. The deep 
learning model can learn complex feature representations from a large number of historical data, and then accurately predict SF6 
gas humidity. 

Keywords: SF6 gas; GIS equipment; Humidity prediction; Deep learning. 
 

1. Introduction 
GIS equipment is one of the key equipment in electric 

power industry, its monitoring and maintenance has been 
concerned. The current research has deeply discussed the 
status and challenges of GIS equipment monitoring and 
maintenance. These studies point out that during the long-
term operation of GIS equipment, the change of SF6 gas 
humidity has an important impact on the performance and 
reliability of the equipment, so humidity prediction becomes 
an important part of GIS equipment monitoring. 

In the field of humidity prediction, a series of traditional 
methods have been applied to GIS equipment SF6 gas 
humidity prediction. These methods include statistics-based 
models, machine learning methods, and physical models. 
However, the traditional method has some limitations in 
processing complex time series data and extracting advanced 
features, which is difficult to meet the needs of practical 
applications. 

In recent years, deep learning, as a powerful data-driven 
method, has gradually emerged in the field of SF6 gas 
humidity prediction. Deep learning models can accurately 
predict the humidity of SF6 gas by learning complex features 
in large-scale data. These models are usually constructed 
based on deep learning algorithms such as convolutional 
neural networks (CNN) and recurrent neural networks (RNN), 
which have high flexibility and predictive performance [1]. 

Although deep learning has made remarkable 
achievements in the field of humidity prediction, its 
application in the GIS equipment SF6 gas humidity prediction 
is still in the preliminary stage. Therefore, it is of great 
significance to deeply study the SF6 gas humidity prediction 
model of GIS equipment based on deep learning, and it is 
expected to provide a new solution for GIS equipment 
monitoring and maintenance. 

2. System Module Function 
(1) Data input and processing module design 

Data input and processing module is an important part of 
SF6 gas humidity prediction model of GIS equipment based 
on deep learning. The module is designed to achieve effective 
acquisition, cleaning and preprocessing of SF6 gas humidity 
data, and provide reliable data support for subsequent model 
training and prediction [2]. 

In the data input phase, the module will obtain SF6 gas 
humidity data from various data sources. These data sources 
can include sensor data in GIS equipment, historical data 
collected by monitoring systems, and external weather 
stations. The data input module with different data sources 
and real-time data acquisition. 

In the data processing stage, the module will clean and 
preprocess the obtained SF6 gas humidity data to improve the 
quality and availability of the data. The cleaning process 
includes removing duplicate data, filling in missing values, 
and handling outliers to ensure data integrity and accuracy. 
The preprocessing involves data normalization, feature 
extraction and data conversion, etc., and the prepared data 
format is input into the deep learning model for training [3-5]. 

(2) The significance and implementation of time step 
alternate setting 

The main purpose of the alternate time step setup is to 
explore the changes in model performance under different 
time steps to better understand and optimize the predictive 
power of the model. 

The significance of time step alternate setting is reflected 
in the model stability and generalization ability, the 
exploration and adaptability of time scale, the optimization 
and adjustment of prediction effect, and the optimization and 
adjustment of prediction effect. 

Therefore, the implementation of an alternate time-step 
setup usually consists of the following steps: 

1) Data division: historical data is divided according to 
different time steps to form training sets, verification sets and 
test sets of different time scales. 

2)Model training and evaluation: Data sets with different 



 

126 

time steps are used to train and evaluate the model. In the 
training process, the model parameters and training strategies 
should be adjusted according to the characteristics of different 
time steps. 

3)Result analysis and comparison: The prediction results of 
the model under different time steps were analyzed and 
compared to evaluate the performance of the model under 
different time scales. Based on the analysis results, the 
optimal time step setting can be selected to optimize the 
prediction effect of the model. 

(3) LSTM network architecture definition and training 
process 

Long Short-term Memory network (LSTM) is a kind of 
deep learning model commonly used for processing sequence 
data, which can effectively capture long-term dependency and 
memory information, and is suitable for modeling and 
forecasting time series data. In the SF6 gas humidity 
prediction model of GIS equipment based on deep learning, 
LSTM network is often selected as the main prediction model, 
and the LSTM network architecture is shown in Figure 1.  

 

 
Figure 1. LSTM network architecture diagram 

 
The LSTM network training process is shown in Figure 2. 

3. Method and Experiment 
(1) Data collection and preprocessing 

1)Data sources and methods of acquisition 

The establishment of SF6 gas humidity prediction model 
requires a large amount of SF6 gas humidity data as the basis 
for training and verification. The data sources mainly include 
GIS equipment sensors, monitoring systems, weather station 
data and so on. Data acquisition methods mainly include real-
time data acquisition, historical data query, and external data 
acquisition. 

2)Data preprocessing procedure 

Before constructing the SF6 gas humidity prediction model 
of GIS equipment based on deep learning, it is necessary to 
preprocess the original data to improve the quality and 
applicability of the data. 

 

 
Figure 2. LSTM network training process 

 

a) Missing value processing 

Missing value is a common problem in data preprocessing, 
which may affect the performance and stability of the model. 
Common missing value handling methods include： 

Delete missing values: In cases where there are fewer 
missing values, you can simply delete the sample or feature 
that contains the missing value. 

Interpolation filling: Using the information of the existing 
data, the missing value is estimated by interpolation methods 
(such as linear interpolation, polynomial interpolation, etc.). 

Fill a constant value: Replaces the missing value with some 
fixed constant value, such as 0 or an average. 

b) Outlier detection and processing 

Outliers may interfere with model training and prediction, 
so outliers need to be detected and processed. Common 
methods include: 

Statistical methods: Statistical methods are used to identify 
outliers in the data. 

Outlier detection algorithm: The outlier detection 
algorithm is used to automatically identify outliers. 

Model-based approach: Outlier detection using machine 
learning models. 

Delete outliers: Delete outliers from data. 

Replace outliers: Replace outliers with the mean, median, 



 

127 

or other appropriate value of the data. 

c) Feature selection 

Feature selection is to select the most relevant and useful 
features from the original features to reduce the complexity 
of the model and improve the prediction performance. 
Common feature selection methods include： 

Filtering method: According to the correlation between the 
feature and the target variable, the commonly used indicators 
include correlation coefficient, Chi-square test and so on. 

Packaging method: Use the performance of the model to 
evaluate the importance of features. Common methods 
include recursive feature elimination. 

Embedding method: Features are automatically selected 
during model training. Common methods include L1 
regularization, decision tree, etc. 

(2) Deep learning model design and training 

1)Model selection 

It is very important to select the appropriate deep learning 
model when constructing the prediction model of SF6 gas 
humidity of GIS equipment based on deep learning. Common 
deep learning models include convolutional neural network 
(CNN) and recurrent neural network (RNN), which each have 
their own characteristics and are suitable for different types of 
data and problems. 

Convolutional neural network (CNN) is mainly used to 
process image data and has good feature extraction ability. In 
the SF6 gas humidity prediction model of GIS equipment, if 
the data may have temporal and spatial characteristics (such 
as spatial relationship and time series) are considered, the SF6 
gas humidity data can be organized into a form similar to 
images, and then processed by CNN. The advantage of CNN 
is that it can automatically extract local features and spatial 
correlations in the data, which is suitable for processing two-
dimensional or three-dimensional data structures. 

Recurrent neural network (RNN) are neural network 
structures specifically designed to process sequence data with 
memory and dynamic modeling capabilities. In the SF6 gas 
humidity prediction model of GIS equipment, SF6 gas 
humidity data are usually collected in time order, which is 
sequential and series-dependent, so RNN is a natural choice. 
With RNN, models can automatically capture temporal 
correlations in data, model historical data, and make future 
predictions. 

2) Model architecture design and optimization 

When designing the architecture of a deep learning model, 
key factors such as the number of layers of the network, the 
number of nodes in each layer, and the choice of activation 
functions need to be taken into account. These factors directly 
affect the representation ability, complexity and prediction 
performance of the model. 

The number of layers of the network determines the depth 
of the model, and deep networks are often able to learn more 
complex feature representations, but are also prone to 
problems such as disappearing gradients or explosions. When 
choosing the number of layers, you need to make trade-offs 
based on the complexity of the data and the difficulty of the 
problem. Typically, you can start with a shallower network 
initially and gradually increase the number of layers until you 
reach a stable point of performance. 

The number of nodes in each layer affects the capacity and 

complexity of the network. The more nodes, the stronger the 
expression ability of the network. However, too many nodes 
can lead to overfitting problems, while too few nodes can lead 
to underfitting. When selecting the number of nodes, we can 
find the appropriate number of nodes to balance the 
performance and generalization ability of the model through 
cross-validation and other methods. 

Activation function plays the role of nonlinear mapping in 
neural network, which can increase the expressive power of 
the network. Common activation functions include ReLU, 
Sigmoid, Tanh, and so on. When selecting the activation 
function, it is necessary to consider the problems such as 
gradient disappearance and gradient explosion, and make 
appropriate selection according to the actual situation. 

In addition to architectural design, regularization and 
optimization techniques need to be considered. 
Regularization methods such as Dropout and L2 
regularization can be used to prevent overfitting, and 
optimization techniques such as Adam and SGD can be used 
to accelerate network convergence and improve training 
efficiency. 

Finally, it is necessary to optimize the hyper parameters of 
the model, including learning rate, batch size, optimizer 
selection, etc. Through cross-validation and other methods, 
the optimal combination of hyper parameters can be found to 
further improve the performance of the model. 

(3)Model performance evaluation and result analysis 

1) Experimental Settings 

When evaluating the performance and analyzing the results 
of SF6 gas humidity prediction model of GIS equipment 
based on deep learning, it is necessary to carry out appropriate 
experimental Settings, including the division of data sets and 
the selection of evaluation indicators. 

Among them, data set partitioning mainly includes training 
set, verification set, test set and so on. The Training Set is used 
to train the model and usually accounts for the majority of the 
total data, which can be around 70% to 80% of the data. 
Validation sets are used for the tuning and parameter selection 
of the model and can account for around 10% to 20% of the 
total data as needed. Test sets are used to ultimately evaluate 
the performance of the model and cannot be used for training 
and tuning of the model, usually accounting for about 10% of 
the total data. The randomness and representations of samples 
should be taken into account in the partitioning of data sets to 
ensure the objectivity and reliability of model evaluation. 

In terms of the selection of evaluation indicators, root mean 
square error (RMSE) is the most commonly used regression 
model evaluation indicator to measure the deviation between 
the predicted value and the actual value. The formula is shown 
in Equation 1. 

 

      RMSE ∑
𝑛

𝑖 1 𝑦 𝑦𝚤      (Equation 1) 

 
Then, the mean absolute error (MAE) is used to measure 

the mean absolute deviation between the predicted value and 
the actual value of the model, as shown in equation 2. 

 

MAE ∑
𝑛

𝑖 1|𝑦 𝑦𝚤|         (Equation 2) 

 



 

128 

Correlation Coefficient is used to measure the degree of 
linear correlation between the predicted value of the model 
and the actual value, and the value range is [-1, 1]. The closer 
the value is to 1, the better the prediction effect is. Finally, 
Goodness of Fit (R-squared) measures the fitting degree of 
the model to the data, and the value range is [0, 1]. The closer 
the value is to 1, the better the fitting effect of the model is. 

2) Model training and evaluation 
During model training and evaluation, data is first prepared, 

loaded according to the previously divided training set, 
verification set and test set, and necessary preprocessing is 
carried out, such as data normalization and standardization. 
Secondly, model initialization initializes the structure and 
parameters of the deep learning model, and selects the 
appropriate optimizer and loss function. The process is then 
trained, the training set is input into the model, and the model 
parameters are constantly adjusted through a back-
propagation algorithm to minimize the loss function. At the 
same time, the validation set is used to verify the model and 
monitor the performance of the model. In the training process, 
the parameters and weights of the model are saved regularly, 
so that the model can be restored and the training can continue 
when the training is interrupted. At the end of the training, 
when the model achieves satisfactory performance on the 

verification set, the training process is stopped and the 
optimal model is saved. 

In the model evaluation, the test set evaluation is carried 
out first, and the trained model is evaluated with the test set. 
The predictive performance indicators of the model on the test 
set, such as RMSE and MAE, are calculated. Secondly, the 
result analysis analyzes the prediction results of the model on 
the test set, compares the difference between the predicted 
value and the real value, discusses the advantages and 
limitations of the model, and further optimizes the model or 
improves the prediction method. Then the visualization 
analysis can be carried out by drawing the prediction curve, 
residual diagram and other ways to visualize the model to 
show the prediction effect of the model. Finally, model 
comparison. If there are multiple models, their performance 
on the same data set can be compared, and the optimal model 
can be selected for practical application. 

3) Result analysis and discussion 
The LSTM network architecture is defined and trained, and 

the single step prediction is realized by using LSTM 
network.The prediction results are compared with the actual 
data. The 300 predictive trainings are shown in Figure 3, and 
the 300 test results since 2000 are shown in Figure 4. 

 

 
Figure 3. Prediction training diagram 

 

 
Figure 4. Test result diagram 



 

129 

 
As can be seen from Figure 4, in the prediction process, 

the model predicted 300 data points after 2000 data points, 
and the data points verified the training results of the neural 
network, showing that the results were good 

4. Conclusion 
Through the construction of GIS equipment SF6 gas 

humidity prediction model based on deep learning, some 
results and enlightenment have been obtained. In the process 
of model design and training, the advantage of deep learning 
algorithm is fully utilized, and the accurate prediction of SF6 
gas humidity is realized through LSTM network and other 
models. In the experiment and result analysis, it is found that 
the model has good performance in the forecasting process, 
and can predict the future humidity level more accurately. 

However, there are still some challenges and shortcomings 
in the construction and training of deep learning models. For 
example, the prediction accuracy of the model is limited by 
the quality and quantity of training data, and the processing 
of complex scenarios and anomalies needs to be further 
improved. In addition, the generalization ability and stability 
of the model also need to be further improved to adapt to the 
actual application requirements in different environments. 

In future studies, it is necessary to continue to optimize the 
design and training methods of deep learning models to 
further improve the performance and stability of predictive 
models. At the same time, more data sources and feature 
representation methods will be explored to improve the 

adaptability and generalization ability of the model. With the 
continuous progress of technology and in-depth research, GIS 
equipment SF6 gas humidity prediction model based on deep 
learning will play an increasingly important role in the 
monitoring and maintenance of power industry. 

References 
[1] Liu Zongyi, Zhang Caihong, Jiang Jianjian, et al. Rapid 

detection of total flavonoids in Dendrobium officinale based on 
Raman spectroscopy combined with CNN-LSTM deep 
learning [J]. Spectroscopy and Spectral Analysis. 2024, 44 (04): 
1018-1024. 

[2] Huang Zehui, Tang Hongbin, Wang Xuesong et al. Research 
on Occupant Injury Prediction Method of Vehicle Collision 
Based on Deep Learning [J]. Automotive Engineer. 2024 
(04):8-11. 

[3] Zou Yutan. Prediction model of soil mechanical properties of 
highway Subgrade based on Machine Learning [J]. 
Transportation Technology and Management, 2019, 5(07):14-
16.  

[4] CHEN Xiaomin, WANG Gang, WANG Longjun. Ordered 
charging strategy of electric vehicle based on vehicle 
networking framework [J]. Modern Electric Power. 2018, 
35(4):1-7. 

[5] Xiaomin Chen, Gang Wang.Optimal Control of Micro Grid 
Operation Mode Seamless Switching Based on Radau 
Allocation Method [C].IOP Conference Series:Materials 
Science and Engineering.2017,199(1):012053