Monitoring of Transformer Hotspot Temperature Using Support Vector Regression Combined with Wireless Mesh Networks

Abstract

The accurate monitoring of the internal hotspot temperature in transformers is crucial for ensuring the stability of power grid operations. Traditional methods typically measure only the surface temperature of transformers, whereas this study proposes a non-invasive thermal inversion algorithm based on a wireless mesh network that effectively predicts the internal hotspot temperature. An electromagnetic-thermal-fluid coupling simulation model was developed to simulate the temperature distribution in transformers under various operating conditions. Subsequently, this study employed a Support Vector Regression (SVR) algorithm to train the sample dataset, optimizing the SVR model using a grid search and cross-validation to enhance the predictive accuracy. After training, the model estimates the hotspot temperature based on surface measurements obtained through a non-contact infrared sensor network. The wireless mesh network, based on the Wi-Fi protocol, provides robust and real-time monitoring even in harsh environments, with data transmitted to a central root node via multiple sensor nodes. The experimental results demonstrate that this method is highly accurate, with predicted temperatures closely matching the results from traditional measurement techniques. This method enhances transformer condition monitoring, helping to extend the transformer lifespan and improve power grid stability.

1. Introduction

Nowadays, power transformers play a very important role in the electrical power distribution grid. The reliability and efficiency of the power distribution grid strongly depend on the safe operation of transformers. The high cost of replacing power transformers, compared to other substation assets, has warranted research aimed at extending their lifespan while ensuring their reliability and stability, primarily through condition monitoring [1]. Transformers are built to last 20 to 30 years and can exceed 40 years if well-maintained. However, damage or aging to any part can decrease their efficiency and lifespan [2]. According to statistical data, a significant number of faults in the power distribution grid were related to faults in transformers [3]. Faults in transformers include many types such as thermal faults, mechanical deformation, dielectric breakdown, and protection faults. Due to the large increase of the electric load in the power distribution grid, thermal faults in power transformers have become the most serious problem [4]. The main reasons for a thermal fault are related to heat generation and dissipation by the transformer. During normal operation, a transformer’s temperature stays stable despite load fluctuations. However, under abnormal conditions such as winding short circuits, the temperature rises above normal levels [5]. If a thermal fault in a transformer is left unattended, it may become worse over time and could lead to a more severe situation. A damaged transformer can lead to widespread power outages. The economic losses from such extensive power failures can also be substantial [6]. To reduce the thermal faults rate in transformers, many approaches have been proposed. Xu developed a roll-core transformer to reduce losses and enhance its air-cooling capacity [7]. Due to the high flash point of the natural ester, its use as an insulator in the transformer’s oil could also enhance the heat endurance of the transformer [8]. Zarko Janic’s study explored the use of variable speed fans to adjust the starting temperature and cooling stage of the fan, which can effectively control the temperature of the transformer, thereby slowing down insulation aging and improving the efficiency and lifespan of the transformer [9]. Although these methods are capable of reducing the rate of thermal faults for the vast majority of the operating transformers, the most effective method in practice is to precisely acquire their temperature. Without effective temperature monitoring and warning, thermal issues will gradually cause thermal faults, which may lead to an emergency accident. Thus, a continuous monitoring system and an operating situational awareness method for power transformers are of great importance for establishing a strong and reliable distribution grid.

In recent years, various types of devices and methods for thermal monitoring have been developed. In 2022, Duan developed a power transformer integrated with distributed optical fiber sensors. The optical fiber sensing principle based on Raman scattering was applied. By laying optical fibers on the windings, the changes in Raman scattering caused by temperature were detected, realizing the continuous monitoring of the winding temperature [10]. In 2023, Abedi et al. used optical fiber sensors (OFSs) and thermography to monitor the temperature of the transformer [11]. Lu developed a real-time temperature monitoring system with distributed optical fiber sensors. These sensors were installed in the transformer cores for measurement purposes [12]. Using gas detection, Ho studied a gas detection system capable of detecting three types of dissolved fault gases in oil-filled power transformers. The gas was used to identify the faults in a power transformer, including the loading and temperature [13]. Despite the intensive work carried out, the sensors, such as thermo-couple and optical fiber, are difficult to set up at the center of the windings because of installation difficulties. Most of these non-contact methods only acquire the surface temperature of the power transformer. Moreover, the measured temperature values are sensitive with respect to the position of the sensors, and the location of the hotspot of the transformer is not easy to determine. In contrast, the internally installed sensors could affect the insulation characteristics of the windings [14]. Therefore, it is crucial to develop a monitoring system, which is harmless to the transformer’s insulation, is easy to deploy, and efficiently monitors the hotspot temperature inside the power transformer without invasion and contact. According to the goal of acquiring the inner hotspot temperature, the main obstacle to obtaining the temperature of the inner hotspot is the difficulty to establish the relationship between the external factors and the inner temperature. Recently presented methods for the thermal calculation are typically based on an empirical formula [15]. The inner temperature of oil-immersed transformers can be calculated as the temperature at the top of the transformer. However, the temperature calculated by the empirical formula is not accurate enough for the increasing demand.

In this paper, we propose a thermal inverse method for monitoring the temperature rise of the inner hotspot of power transformers, based on a wireless mesh network. The primary objective is to develop a non-invasive, real-time monitoring system that accurately determines the inner hotspot temperature. To achieve this, we first introduce the design of the monitoring mesh network in Section 2, covering the system requirements, an overview of the system, and the measurement procedure. In Section 3, we present the simulations and experimental verification of the power transformer, including the development of an electromagnetic-thermal-fluid-coupled simulation model to analyze the heat distribution and identify hotspots, followed by validation through temperature rise experiments. Section 4 focuses on the development of the thermal inverse method using Support Vector Regression (SVR), detailing the establishment of the SVR model, construction of the sample set, and evaluation of the inversion model. In Section 5, we discuss the advantages and limitations of our method, the impact of key parameters, and the role of Wireless Mesh Networks in enhancing the system’s effectiveness. Finally, Section 6 concludes the paper, summarizing the key findings and emphasizing the significance of the proposed method for transformer temperature monitoring and the operation of power grids.

2. The Design of the Monitoring Mesh Network

2.1. Requirements of the Monitoring System

To avoid damage to a transformer’s insulation characteristics caused by wired sensors, a thermal monitoring system should not contact the surface or be placed too close to the windings of the transformer. Moreover, the sensors’ response time should be sufficiently short to meet the requirement for real-time monitoring. Compared with previous approaches, the proposed monitoring system must provide a flexible arrangement and easy installation. Considering the operating temperature of a power transformer, the temperature measurement range should be from −40 to 300 degrees centigrade.

2.2. Brief Introduction of the Proposed Monitoring System

As illustrated in Figure 1, the wireless system for thermal monitoring is based on a mesh network. The mesh network is built atop the Wi-Fi protocol. The network allows nodes to spread around a transformer and be interconnected in a single WLAN. The mesh network is self-organizing and self-healing and can be built and maintained autonomously.

The root node is the top-level node in the mesh network, serving as the sole interface between the mesh network and the external network. Sensor nodes are placed around the power transformers at a considerable distance from the root node, transmitting measurement information to the root node through one or more wireless hops. Due to the design requirements of this section for a non-invasive online temperature acquisition sensor network, a non-contact infrared temperature sensor is used to collect the surface temperature data of operating transformers. However, since the temperature sensor itself does not have wireless communication capabilities, it requires an external ESP32 communication chip (Espressif Systems, Wuxi, Jiangsu, China). This chip forms a Wi-Fi Mesh network to transmit temperature data to the root node in the wireless sensor network, which will ultimately be used by clients in the external network.

As the temperature sensor outputs temperature data in an RS232 or RS485 communication signal format, and the ESP32 communication chip can only process TTL-level signals, a serial converter is needed to convert between the two signal types to ensure normal communication between the wireless sensor network and the external network. The physical diagram of the temperature sensor node in the non-invasive online temperature acquisition wireless sensor network is shown in Figure 2.

2.3. Measurement Procedure of the Monitoring System

The measurement procedure of the proposed monitoring system is illustrated in Figure 3. Based on the thermal monitoring system, the surface temperature values of each phase are captured by the infrared nodes, which transmit these values to the root node. Then, the inner hotspot temperature of the transformer can be calculated by the root node, whereas the transformer is in operation.

In the proposed system, users can use their remote device to obtain the temperature of the inner hotspot of the transformer through a web page generated by the root node. If the temperature is above a warning level, a message will be displayed on the web page.

3. Simulations of the Power Transformer and Experimental Verification

3.1. Calculation of the Thermal Distribution and Hotspot

A typical power transformer is shown in Figure 4. The transformer windings are made of aluminum wire, and the insulating layer is formed on the aluminum surface by the oxidation process. The rated power of the three-phase transformer is 100 kVA, and its rated voltage is 315/400 V.

To determine the inner hotspot of the power transformer, an electromagnetic-thermal-flow-coupled simulation model is established.

As illustrated in Figure 5, the 3D model of the transformer is built. The typical parameters of the transformer are provided in

Table 1. Transformer Parameters.

Parameters	Primary Windings	Secondary Windings
Turns	105	78
Connection	Triangle	Star with neutral
Rated winding currents	105.8 A	144.3 A
Wire cross-sections	10 × 5 mm	15 × 5 mm

.

The heat sources in the power transformer originate mainly from iron losses and winding losses. To simulate the heat distribution of the transformer, these losses should be studied first. As described in Equation (1), the iron losses are composed of the eddy current losses and hysteresis losses [16]. The eddy current losses are produced by AC in the windings, and the hysteresis losses are determined by the hysteresis characteristics of the core material.

$$P_{core}=P_h+P_e=k_{h}f{B}^{\partial }+k_e{\left(fB\right)}^2$$

(1)

where P_core is the iron loss, P_h is the hysteresis loss, P_e is the eddy current loss, k_h is the hysteresis loss coefficient, k_e is the eddy current loss coefficient, ∂ is the Steinmetz coefficient, f is the frequency of the magnetic field, and B is the magnetic flux density.

According to the B–H curve of the iron core, the calculation of the magnetic flux density is based on the finite element method (FEM) for the short-circuit condition.

Winding losses can be divided into eddy current losses and wire losses. The wire losses can be calculated by Equation (2).

$$Q={\displaystyle \frac{I^{2}R}{V}}$$

(2)

where Q is the wire loss per unit volume, R is the resistance of the windings, I is the current, and V is the unit volume.

Due to the cross-section of the windings, the wire losses are relatively low. However, the eddy current losses of the aluminum wire cannot be neglected. The FEM calculation of the eddy current loss is based on the following equation:

$$P_{e_{-}mine}={\displaystyle \frac{1}{24}}\sigma w^2{d}^2{B}^2$$

(3)

where σ is the conductivity of the windings, ω is the angular frequency, d is the thickness of each wire, and B is the magnetic flux density.

Based on the calculation of losses, the heat sources can be determined and imported to the heat simulations. The heat distribution in the transformer is mainly related to heat conduction and convection. According to Equation (4), the heat generated by the iron core and windings can be conducted from the internal to the external parts of the transformer and other metal parts [17].

$$\rho C_p{\displaystyle \frac{\partial T}{\partial t}}+\rho C_{p}u\cdot \nabla T+\nabla \cdot (-k\nabla T)=Q$$

(4)

where ρ is the density of the material, C_p is the specific heat at a constant pressure, T is the temperature, u is the flow rate, k is the thermal conductivity, and Q is the heat source.

In consideration of the heat convection, the heat can be dissipated to the air by natural convection because of the structure of the transformer. This is described by Equation (5).

$$Q=hS\left(T_w-T\right)$$

(5)

where h is the convective heat transfer coefficient, T_w is the environment temperature, and S is the selected area.

The temperature distribution of the transformers under different operating conditions was simulated using finite element software, and the simulation results were verified by building a temperature rise experimental platform. The temperature distribution of the transformer is simulated under a no-load operation, ambient temperature of 293.15 K, and no wind conditions, as shown in Figure 6.

The transformer operates under three-phase unbalanced conditions, with an ambient temperature of 293.15 K and no wind. The temperature distribution of each component of the transformer is simulated and shown in Figure 7.

By comparing Figure 6 and Figure 7, it can be seen that during the operation of the transformer, the intersection of the middle winding and the upper iron yoke is the hottest area of the transformer, and the hotspot is generally located in this area. However, when the transformer operates under a three-phase unbalanced load, the rightmost winding has the highest load current. It can be seen that the hottest position of the transformer moves towards the winding with the highest current. Therefore, it can be determined that the middle phase and upper iron yoke of the transformer are the hottest areas during operation, and the hottest position is related to the magnitude of the load current of each phase of the transformer winding, that is, the hottest position of the transformer shifts towards the winding with the highest load current.

3.2. Experimental Verification

To validate the accuracy of the transformer simulation model proposed in this paper, a temperature rise experiment was conducted. The purpose of the experiment was to verify the accuracy of the internal hotspot temperature calculated by the simulation model by measuring the surface temperature of the transformer under different load conditions. The experiment was conducted under laboratory conditions using a three-phase transformer with a rated power of 100 kVA.

The experiment was conducted under no-load, 0.3-load, and three-phase unbalanced load conditions. An infrared thermal imaging camera was used to measure the surface temperature distribution of the transformer, with the measurement distance and angle kept constant to ensure data accuracy. Data were recorded after the transformer’s temperature stabilized under each load condition.

Under no-load operation conditions, the infrared thermal images of the steady-state surface temperature field of the transformer are shown in Figure 8.

Based on the existing conditions in the laboratory, a transformer temperature rise experimental platform was constructed using a forced air-cooled low-voltage load cabinet connected to the high-voltage side winding of the transformer, as shown in Figure 9.

According to the definition of the load factor, β = I₂/I_2N, where I₂ is the output current of the transformer and I_2N is the rated output current of the transformer. The calculation shows that the transformer is operating at a load rate of 0.3 times. Under this operating condition, the infrared thermal image of the steady-state surface temperature field of the transformer is shown in Figure 10.

When a transformer operates under three-phase unbalanced conditions, it outputs a three-phase asymmetric current. The three-phase unbalanced current will cause a change in the distribution of internal heat sources in the transformer, resulting in a different temperature distribution of the transformer compared to the temperature distribution under different load rates, as shown in Figure 11.

The main sources of error in this experiment include sensor accuracy errors, environmental factors, and measurement position deviations. The Fluke TiS10 infrared thermometer (Fluke Corporation, Everett, WA, USA) used in this study, has limitations in accuracy, and discrepancies between the measured and true values are unavoidable. In addition, environmental factors such as temperature, humidity, and wind speed can affect the transformer’s heat dissipation. To minimize these effects, the experiment was conducted over several consecutive days under stable environmental conditions at the same time intervals. Finally, the measurement position deviations were reduced by taking multiple measurements at the same spot.

By comparing the temperature rise experimental thermal imaging of aluminum winding ceramic insulated transformers under three-phase symmetrical working conditions, it was found that the temperature of the intermediate-phase winding was significantly higher than that of the other two-phase windings, and the upper temperature was higher than the lower temperature, with the highest temperature occurring at 90% of the height of the intermediate-phase winding. The thermal image of the temperature rise experiment is consistent with the simulation results of the temperature fluid coupling model, which verifies the accuracy of the multi-physics field coupling model in this paper and provides accurate data support for the research of the winding hot spot inversion method in the following text.

4. Thermal Inverse Method for the Evaluation of the Inner Hotspot Temperature of a Power Transformer

4.1. Establishment of SVR

This research focuses on modeling the nonlinear relationship between multiple input parameters and a single output. Common models include neural networks, decision tree regression, and SVR. Neural networks can capture complex relationships, but require large datasets, are computationally intensive, and risk overfitting with small samples. Decision tree regression divides the input space into regions based on attribute thresholds and makes predictions within each region. However, it struggles with continuous data and is sensitive to noise.

This study uses SVR, which offers significant advantages. SVR is widely used in pattern recognition and function fitting [18,19,20]. SVR is based on structural risk minimization, balancing model complexity and fitting accuracy by finding the optimal decision boundary through support vectors. It is effective even with small sample sizes and provides accurate predictions of the hotspot temperature, making it a strong choice for this study. The basic principle is shown in Figure 12.

During the implementation of the SVR algorithm, we undertook the following specific steps. First, in terms of parameter selection and optimization, the key parameters of the SVR model include the penalty coefficient C and the kernel function parameter γ.

The penalty coefficient (C) controls the trade-off between the model complexity and training error. Based on prior studies and initial tests, the search range for C was set to [0.1, 100]. A small C value simplifies the model, but increases the training error, leading to underfitting, while a large C value may cause overfitting. A step size of 0.5 was chosen to balance the thorough exploration of the parameter space with computational efficiency. By adjusting C in increments of 0.5, the model’s accuracy and generalization ability were analyzed.

The kernel parameter (γ) influences how input features are mapped to a high-dimensional space. Given the transformer temperature data’s complex, nonlinear relationships, the search range for γ was set to [0.001, 10]. A small γ value leads to a smoother decision boundary, potentially causing underfitting, while a large γ value can lead to overfitting. A step size of 0.01 was used to explore the effect of different γ values on the model performance in detail.

To optimize these parameters, we employed a Grid Search algorithm combined with K-Fold Cross-Validation. During the grid search process, we formed a parameter combination matrix according to the set parameter search ranges and step sizes as described above. Then, K-Fold cross-validation was performed for each parameter combination, and finally, the optimal parameter combination was selected.

In the training process, the sample data were first standardized to ensure that all feature values were on the same scale for training. Then, the optimized SVR model was used to learn from the training set, constructing a regression model. By training the model with temperature data under different working conditions, accurate predictions of the internal hotspot temperature of the transformer were achieved.

For the training sample set {(x_i, y_i)| i = 1, 2, …, N, x_i ∈ R_N, y_i ∈ R}, by introducing the insensitive loss coefficient ϵ, the distance between the predicted value f(x_i) of the decision hyperplane f(x) and the true value y_i of the input vector x_i is minimized, which can be mathematically expressed as shown in Formula (6).

|y_i − f(x_i)| ≤ ε, i = 1, 2, …, N

(6)

4.2. Establishment of Sample Set

Constructing an accurate hotspot temperature inversion model heavily relies on the construction of the sample set. This section constructs the sample set using the transformer temperature field simulation results under different operating conditions as the source of sample data. From the temperature field diagrams in Figure 8, Figure 9, Figure 10 and Figure 11, it can be observed that the hotspots of the transformer windings are generally located at approximately 90% of the height of the middle-phase winding. Based on the analysis of the transformer’s heat dissipation characteristics, this paper selects the locations shown in Figure 13 as the surface temperature measurement points.

Based on the concept of Principal Component Analysis (PCA), the characteristic temperature measurement points that have a strong correlation between the surface of the transformer and the winding hot spots are selected from the initially chosen surface temperature measurement points shown in Figure 13. The main approach is illustrated in Figure 14.

The purpose of Principal Component Analysis (PCA) is to find a new set of variables that can replace the original variables. These new variables must be uncorrelated with each other and be able to carry the maximum amount of useful information from the original variables. In the PCA algorithm, eigenvalues represent the extent to which their corresponding principal components contribute to the original data matrix. Similarly, the contribution rate can more intuitively represent the extent to which the principal components retain the original information. As shown in the scree plot in Figure 15, the cumulative contribution rate of the first three principal components is 95.768%. Therefore, this paper selects the first three principal components for analysis.

By calculating the component score coefficient matrix and analyzing the relationship between each temperature measurement point and the extracted common factors, and considering the influence of the non-isothermal fluid flow on the transformer’s surface temperature, this paper selects measurement points 3, 4, 9, 10, 13, and 14 as the characteristic temperature measurement points for the inversion of the transformer winding hotspots.

Since the temperature change of the transformer is a relatively slow process, the temperature field at each moment can be approximately considered as a quasi-static field. In the process of constructing the sample set, to increase the number of samples, the temperature at the characteristic measurement points from the temperature field simulation results at different times under various operating conditions was recorded with a certain step size. In this paper, a transient simulation analysis of the transformer’s temperature field was performed using finite element software with a time step of 0.15 h and a simulation duration of 12 h. The conditions included no-load operation, rated-load operation, 0.3-, 0.6-, 0.9-, 1.1-, and 1.2-times load ratios, and three-phase unbalanced load operation. Combined with the three-phase load ratio as the input load features, a sample set for winding hotspot inversion was established, as shown in

Table 2. Winding Hotspot Temperature Inversion Sample Set (Excerpt).

Number	ALR	BLR	CLR	3#	4#	9#	10#	13#	14#	Hs
1	0.3	0.3	0.3	89.2	109.8	73.2	64.6	80.6	100.2	128.6
2	0.3	0.3	0.3	113.2	82.6	75.6	74.3	92.1	86.5	135.2
3	0	0	0	73.2	84.6	64.8	72.7	71.8	78.2	95.2
4	0	0	0	85.8	91.4	71.2	75.0	74.2	79.5	109.2
5	0.5	0.6	0.65	88.3	82.3	69.8	75.2	67.7	78.8	128.6
6	0.5	0.6	0.65	106.3	93.3	84.3	86.4	75.2	89.8	144.7

. In the table, ALR, BLR, and CLR represent the A-phase load rate, B-phase load rate, and C-phase load rate, respectively.

4.3. Transformer Winding Hotspot Temperature Inversion Detection

Using multi-dimensional feature quantities as inputs and the transformer winding hotspot temperature as the output, a grid search method was employed to exhaustively search the hyperparameter combinations of the SVR model. A 5-fold cross-validation was then used to identify the hyperparameter combination that optimizes the model’s performance. The best-performing model was selected as the winding hotspot inversion detection model. The flowchart of the SVR algorithm based on grid search tuning and 5-fold cross-validation is shown in Figure 16.

The improvements in grid search optimization and cross-validation algorithms allow the inversion model to achieve high accuracy even with a small sample set. The final optimized penalty coefficient and kernel function parameters are shown in

Table 3. Optimization results of the grid search algorithm.

Parameter	Result
$C$	22.6274
$\gamma$	1.0215

.

In Table 3, C represents the penalty coefficient, which determines the generalization ability of the model, and γ represents the kernel function parameter, which determines the effectiveness of the linear separability of the data after nonlinear mapping.

To analyze the effectiveness of transformer hotspot inversion detection based on feature temperature measurement points, three error evaluation metrics were used to assess the transformer hotspot inversion model, as shown in

Table 4. Inversion error metrics for feature temperature measurement points.

Evaluating Indicator	Train Set			Test Set
	${\mathit{e}}_{\mathit{M}\mathit{A}\mathit{E}}$	${\mathit{e}}_{\mathit{M}\mathit{A}\mathit{P}\mathit{E}}$	${\mathit{e}}_{\mathit{R}\mathit{M}\mathit{S}\mathit{E}}$	${\mathit{e}}_{\mathit{M}\mathit{A}\mathit{E}}$	${\mathit{e}}_{\mathit{M}\mathit{A}\mathit{P}\mathit{E}}$	${\mathit{e}}_{\mathit{R}\mathit{M}\mathit{S}\mathit{E}}$
Results	3.3464	2.5926	4.4764	5.6663	5.1723	6.1354

.

In Table 4, $e_{MAE}$ represents the mean absolute error, $e_{MAPE}$ represents the mean absolute percentage error, and $e_{RMSE}$ represents the root mean square error.

4.4. Verification of Inversion Detection

To verify the thermal inverse method proposed in this study, an experiment was carried out between the optical fiber and the monitoring mesh network. During the experiment, the sensor nodes with an infrared detector transmitted the thermal information to the root node, as well as the node containing the sensor, which monitors the environment temperature.

According to the finite element simulation results of the transformer temperature field, the error between the platform-detected transformer winding hotspot temperature and the actual winding hotspot value was 2.5 °C. Compared with the calculation method of the thermal network model mentioned in reference [21], the accuracy of the method proposed in this study has been improved by about 3.26%.

5. Discussion

As the research in transformer temperature monitoring progresses, it is essential to evaluate and compare different methods to understand their strengths and limitations. In this regard, the following discussion delves into the details of various approaches and their implications.

In recent years, several temperature inversion methods have been proposed. For instance, Sun et al. developed an online temperature monitoring system integrating a Pt100 sensor and an infrared thermal imager. However, the installation of the Pt100 sensor is inconvenient and may impact the transformer’s interior, and the system’s flexibility is limited [22]. Hao et al. put forward a more accurate inversion algorithm, yet they did not provide an online monitoring solution [23]. In contrast, this study presents a more efficient and adaptable online monitoring approach. It overcomes the drawbacks of previous methods by avoiding invasive sensors and enhancing system flexibility and scalability. Nevertheless, this method also has its limitations. The complexity of real-world operating conditions, especially in extreme environments or under sudden load variations beyond the model’s learning scope (like high-current surges due to lightning strikes or short-circuit faults), can pose challenges to the prediction accuracy. Additionally, transformer aging, which alters thermal characteristics, may lead to decreased prediction accuracy if the model is not updated.

Key parameters such as the load rate, the position, and the number of surface temperature measurement points are crucial to the prediction accuracy. The load rate impacts heat generation and distribution, and under three-phase unbalanced conditions, it significantly affects the heat source distribution. The arrangement of the temperature measurement points, constrained by the transformer structure, can also affect the accuracy.

Wireless Mesh Networks present notable benefits in the context of monitoring large transformers or multiple units within substations. They bring cost savings, enable non-intrusive installation, possess scalability, and allow for flexible system upgrades. Their self-organizing and self-healing capabilities enhance the communication stability, facilitating the real-time and centralized monitoring of transformer temperatures. However, in the complex electromagnetic environment of substations, electromagnetic interference is a significant challenge as it can disrupt wireless communication, resulting in data loss or delays. To address this issue, measures such as network topology optimization, redundancy implementation, electromagnetic shielding, and the utilization of more interference-resistant protocols and frequencies can be employed to enhance the reliability of the system.

6. Conclusions

This article proposes an online monitoring method for the transformer winding hotspot temperature, from which the following four conclusions can be drawn:

(1)

This study presents a wireless mesh network-based thermal monitoring system using SVR. Through simulation and experimental data, it shows an excellent predictive performance for transformer hotspot temperatures. The simulation considered various factors like different load levels and ambient temperatures, and the experimental validation on real transformers further verified its accuracy. This combination of wireless technology and SVR enables a precise and non-invasive prediction, making it a valuable approach for transformer temperature monitoring.

(2)

The experimental results confirm the effectiveness of the SVR method. The predicted hotspot temperatures closely match traditional measurements. In terms of MAE, MAPE, and RMSE, it outperforms other methods, demonstrating strong fitting and generalization capabilities. This means the method can accurately model the relationship between input variables and the hotspot temperature, which is crucial for practical use in transformer operation.

(3)

The developed system is non-invasive, real-time, and stable. Using non-contact infrared sensors, it avoids physical intrusion into the transformer. The real-time monitoring allows for the immediate detection of temperature changes, and its stability ensures reliable operation even in complex and harsh environments like substations. This system plays a vital role in ensuring the safe operation of the transformer by providing accurate and timely temperature information.

(4)

The method is adaptable and scalable. The wireless mesh network allows for easy expansion and modification, laying the groundwork for multi-transformer monitoring in power grids. It has broad application potential and supports the intelligent development of power grids by enabling more efficient temperature monitoring.

Future research could focus on applying this method to different types and sizes of transformers, so as to enhance the universality of this method and better meet the diverse needs of the power industry. Additionally, it is also necessary to conduct further research on how to optimize the system to achieve long-term performance and reliability.