Dynamic prediction method and system for hot spot temperature of transformer under alternating load

Abstract

The application provides a transformer hotspot temperature dynamic prediction method and system under alternating load, training data sets are constructed by fusing different frequency alternating load simulation data generated by CFD simulation and transformer historical operation data; a TCN-Attention hybrid model is designed, long sequence fluctuation characteristics are extracted by using causal expansion convolution, and key feature points such as load wave peak and wave trough are focused by combining an attention mechanism; a transformer heat conduction equation is embedded into a model loss function as a physical constraint, a back propagation algorithm of deep reinforcement learning is used for training, and a hotspot temperature prediction model is obtained; based on real-time collected data, a future short time hotspot temperature prediction trajectory is output, and an insulation paper thermal aging coefficient is calculated to generate a graded early warning signal. The application realizes deep fusion of physical mechanism and data driving, solves the technical problems of low prediction accuracy under alternating load caused by wind and light fluctuation and divergence of extreme working conditions, and significantly improves the robustness and explainability of the prediction trajectory.

Description

Technical Field

[0001] This invention relates to the field of power transformer condition monitoring, and in particular to a method and system for dynamic prediction of transformer hot spot temperature under alternating load. It is applicable to real-time dynamic prediction of power transformer winding hot spot temperature and insulation life assessment under alternating load conditions in grid-connected scenarios of new energy sources such as wind power, photovoltaic, and wind-solar-storage integration. Background Technology

[0002] With the rapid development of the new energy industry, distributed energy sources such as wind power and photovoltaics are being connected to the grid on a large scale. Due to the inherent randomness, intermittency, and volatility of wind and solar power output, power transformers connected to the grid are operating under alternating load conditions for extended periods, resulting in load curves exhibiting typical "sawtooth" drastic fluctuations. The hot spot temperature of the transformer windings is a core indicator determining the aging rate of its insulation paper and affecting the equipment's service life. Excessively high hot spot temperatures directly accelerate insulation failure and can even lead to serious faults such as winding breakdown and transformer burnout. Therefore, accurate real-time prediction of hot spot temperatures under alternating loads is of great significance for transformer condition monitoring and operation and maintenance.

[0003] Currently, transformer hot spot temperature prediction technology is mainly divided into two categories: traditional thermal circuit models and pure data-driven models. 1. Traditional thermal circuit models, represented by the thermoelectric analogy method recommended by the IEEE C57.91 standard, simplify the transformer into a first- or second-order thermal circuit and calculate the hot spot temperature through the thermoelectric analogy relationship between heat capacity and thermal resistance. However, this type of model is based on steady-state or slowly changing load assumptions, has a large time constant, and significant inertial errors. It cannot track the drastic load changes on the order of seconds or minutes under alternating loads, and the prediction results have a significant lag, making it difficult to adapt to alternating load conditions caused by wind and solar fluctuations.

[0004] 2. Pure data-driven models, such as deep learning models like LSTM and GRU, achieve temperature prediction through training with a large amount of historical data. They have high accuracy under stable loads. However, these models are "black box" models, which do not incorporate the physical mechanism of heat conduction in transformers. Under extreme conditions (such as sudden increases or decreases in load, missing data, or overload operation), the prediction results are prone to divergence and violation of physical laws. For example, the predicted temperature may rise indefinitely, making the prediction results unreliable.

[0005] In addition, some existing solutions attempt to combine CFD simulation with machine learning models, but they do not optimize the network structure for the long-sequence fluctuation characteristics of alternating loads, nor do they embed the heat conduction equation as a strong physical constraint into the model training. This results in the model's insufficient ability to capture the features of alternating loads, and its generalization ability under extreme conditions still needs to be improved.

[0006] Furthermore, transformer hotspot temperature prediction technologies are mainly divided into two categories: traditional thermal circuit models and pure data-driven models. Traditional thermal circuit models, represented by the thermoelectric analogy method recommended by relevant standards, simplify the transformer into a first- or second-order thermal circuit for temperature prediction. For example, the existing technology CN121328101A proposes a method and device for evaluating and analyzing heavy overload of transformers. It establishes a thermal circuit model based on the heat generation power, equivalent heat capacity, and equivalent thermal resistance of the transformer's internal windings, and calculates the temperature by transforming it into an equivalent heat pressure source based on Thevenin's theorem. However, these models are mainly based on steady-state or slowly changing load assumptions, have large time constants, and significant inertial errors. They cannot track the drastic load changes on a second or minute scale under alternating loads, and the prediction results have significant lag, making it difficult to adapt to alternating load conditions caused by wind and solar fluctuations.

[0007] On the other hand, pure data-driven deep learning models achieve temperature prediction through training on a large amount of historical data. For example, the existing technology CN118862496A discloses a method for determining the hot spot temperature of a transformer. It obtains the temperature distribution data of the transformer under different environments by constructing a fluid-thermal field coupling model, and then trains a recurrent neural network model to predict the hot spot temperature of the transformer at various times. Although such methods have a certain accuracy under stable loads, because pure data-driven models are "black box" models and do not deeply integrate the underlying physical mechanisms of transformer heat conduction, they are prone to diverging prediction results and violating physical laws when faced with extreme operating conditions such as sudden increases or decreases in load, data gaps, and overload operation, resulting in unreliable prediction results.

[0008] Furthermore, some existing technologies attempt to improve prediction and protection capabilities by addressing external factors such as the cooling system status. For example, the existing technology CN120893306B proposes a transformer overheat protection method and system, which obtains a cooling activation index by analyzing the operating power of the cooling system, generates a cooling status label, and integrates it into the backbone network structure for temperature prediction. However, these existing solutions still fail to deeply optimize the network extraction structure for the long-sequence fluctuation characteristics of wind and solar alternating loads, nor do they directly embed rigorous heat conduction physical equations as strong physical constraints into the loss function of model training. This results in insufficient feature capture capability of the model for complex alternating loads, and the generalization ability and reliability under extreme conditions still need further improvement.

[0009] Therefore, there is an urgent need to develop a transformer hotspot temperature prediction technology that optimizes for alternating load characteristics and integrates the advantages of physical mechanisms and data-driven approaches. This technology would address the problems of large inertial errors in traditional models and the failure of purely data-driven models under extreme conditions, achieving high-precision and robust dynamic prediction of hotspot temperatures under alternating loads. (Note: The last sentence appears to be unrelated and possibly a fragment from another context: "damaging lens surface.") Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for dynamic prediction of transformer hotspot temperature under alternating load based on deep reinforcement learning. Targeting the alternating load conditions of transformers caused by wind and solar fluctuations, this invention utilizes CFD simulation data augmentation, TCN-Attention model optimization, and physical constraint embedding techniques to combine the strong constraints of physical mechanisms with the high fitting capability of data-driven approaches. This achieves real-time and accurate dynamic prediction of transformer hotspot temperature, solving the technical problems of large inertial errors in traditional thermal circuit models and failure of purely data-driven models under extreme conditions. Simultaneously, it provides a basis for transformer insulation life assessment and graded early warning.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for dynamic prediction of transformer hot spot temperature under alternating load, the method comprising: S1. Construct a training dataset under alternating transformer load; S2. Design a Temporal Convolutional Network Attention Mechanism (TCN-Attention Hybrid Model); S3. Embed the transformer heat conduction equation as a physical constraint into the loss function of the TCN-Attention hybrid model to train a hotspot temperature prediction model; S4. Input the real-time alternating load data of the transformer into the trained hot spot temperature prediction model, and output the short-term predicted trajectory of the transformer winding hot spot temperature. S5. Calculate the thermal aging coefficient of transformer insulation paper based on the hot spot temperature prediction trajectory.

[0012] In the preferred embodiment, step S1 involves constructing a training dataset under alternating transformer load, and the steps include: 1a. Establish a three-dimensional geometric model of the transformer winding, oil passage and insulation structure, set the alternating load waveforms of different frequencies as simulation boundary conditions, solve the oil flow field and temperature field distribution through CFD simulation, and extract the temperature time series data of the hot spot location of the winding. The training dataset contains transformer oil flow field and temperature field data under alternating loads at different frequencies generated by CFD simulation, as well as historical transformer operating data; 1b. Collect historical operating data of the transformer, which includes at least the load current sequence, ambient temperature sequence, top oil temperature sequence, and measured hot spot temperature sequence of the winding; 1c. The CFD simulation data and historical running data are preprocessed, fused, and divided into training set, validation set, and test set to obtain the training dataset.

[0013] In the preferred embodiment, the data preprocessing in step 1c includes wavelet transform denoising, feature extraction, data normalization, and outlier removal performed sequentially. The extracted features include the high-frequency fluctuation component of the load current, the rate of change of current, and the temperature gradient features.

[0014] In the preferred scheme, step S2 involves designing the TCN-Attention hybrid model, which includes the following steps: 2a. Construct a TCN feature extraction layer, using stacked causal dilated convolution kernels to form a multi-level network structure, extract the time-dependent features of long alternating load sequences, and solve the gradient vanishing problem through residual connections; 2b. Connect an attention mechanism layer to the output of the TCN feature extraction layer to calculate the weight distribution of load features at different time steps and focus on the key feature points of the alternating load peaks and troughs. 2c. Construct the model output layer and map the output features of the attention mechanism layer to the predicted values ​​of transformer hot spot temperature.

[0015] In the preferred embodiment, step S3 embeds the transformer heat conduction equation as a physical constraint into the loss function. The steps include: 5a. Based on the heat capacity-thermal resistance model of the transformer oil-paper insulation system, the heat conduction equation is determined as follows: ,in For equivalent heat capacity, For equivalent thermal resistance, For hot spot temperature, For ambient temperature, The loss source term varies with the square of the load current; 5b. Construct the joint loss function ,in The mean square error between the model's predicted values ​​and the measured values. This represents the sum of squared residuals of the heat conduction equation in a discrete grid. This is a balancing coefficient used to adjust the weights of data fitting and physical constraints. 5c. The backpropagation algorithm of deep reinforcement learning is used in combination with the Adam optimizer to train the model until the validation set loss converges.

[0016] In the preferred scheme, the short time in S4 is 5min~72h. The real-time alternating load data of the transformer includes the real-time collected three-phase load current, ambient temperature and top oil temperature. Before inputting it into the model, it needs to be preprocessed in the same way as in S1.

[0017] In the preferred scheme, the thermal aging coefficient is calculated in S5 using the six-degree rule or the Montsinger chemical kinetic equation, specifically as follows: Based on the predicted hot spot temperature trajectory, the aging rate of the insulating paper at different time steps is calculated, and the cumulative thermal aging coefficient is obtained by integrating over time. When the predicted hot spot temperature exceeds the preset temperature threshold, or the cumulative thermal aging coefficient exceeds the preset aging threshold, a graded early warning signal for transformer overheating or life loss is generated.

[0018] In the preferred embodiment, the method is deployed in the edge computing gateway of the new energy power plant, and the real-time dynamic prediction of transformer hot spot temperature is achieved through edge computing; New energy power plants are wind power, photovoltaic power, or integrated wind-solar-storage power plants.

[0019] A dynamic prediction system for transformer hotspot temperature under alternating load based on deep reinforcement learning includes: The data construction module is used to build a training dataset for transformers under alternating loads, integrate CFD simulation data and historical transformer operating data, and complete data preprocessing. The model design module is used to design the TCN-Attention hybrid model and build a network architecture that includes a TCN feature extraction layer, an attention mechanism layer, and an output layer. The model training module is used to embed the transformer heat conduction equation as a physical constraint into the model loss function, and to train the hot spot temperature prediction model using a deep reinforcement learning algorithm. The prediction output module is used to receive real-time alternating load data of the transformer, input it into the trained prediction model, and output the short-term predicted trajectory of the hot spot temperature. The aging assessment module is used to calculate the thermal aging coefficient based on the hot spot temperature prediction trajectory and generate graded early warning signals according to preset thresholds.

[0020] In the preferred embodiment, the CFD simulation unit is used to establish a three-dimensional geometric model of the transformer, set boundary conditions for alternating loads at different frequencies, and simulate and generate oil flow field and temperature field data. The data acquisition unit is used to collect historical operating data and real-time alternating load data of the transformer through sensors; The data preprocessing unit is used to perform wavelet transform denoising, feature extraction, normalization, and outlier removal on simulation data and measured data; The model design module includes a TCN building unit and an attention mechanism building unit; The TCN building block is used to build a multi-level TCN network with causal dilated convolution, and to set residual connections and dilation factors. The attention mechanism building unit is used to design the weight calculation module and generate attention weight matrices for load features at different time steps; The model training module includes a physical constraint embedding unit and a training optimization unit; The physical constraint embedding unit is used to determine the heat conduction equation based on the transformer thermal capacity-thermal resistance model and construct a joint loss function that includes data fitting terms and physical residual terms. The training and optimization unit is used to train the model using the deep reinforcement learning backpropagation algorithm combined with the Adam optimizer, thereby updating and optimizing the network weights.

[0021] The prediction output module adopts a sequence-to-sequence encoding-decoding structure. The encoding end extracts the temporal features of the real-time load data, and the decoding end outputs the predicted hotspot temperatures at different time steps in the future, forming a continuous temperature prediction trajectory.

[0022] This invention provides a method and system for dynamically predicting the hot spot temperature of a transformer under alternating load. Compared with the prior art, this invention has the following significant advantages: 1. High prediction accuracy and small inertial error: The TCN-Attention hybrid model is designed for the long-sequence fluctuation characteristics of alternating loads. Causal dilated convolution can effectively capture the long-period dependence characteristics of the load, and the attention mechanism focuses on key feature points, which solves the inertial error problem of traditional thermal circuit models and realizes accurate dynamic prediction of hot spot temperature under alternating loads. The measured error can be controlled within 3℃.

[0023] 2. Strong robustness and effectiveness under extreme conditions: By embedding the heat conduction equation as a physical constraint into the loss function, the model prediction results always converge and conform to the physical laws of heat conduction. This solves the "black box" failure problem of pure data-driven models under extreme conditions. Even under unseen conditions such as sudden load increases and missing data, it can still output reliable prediction results.

[0024] 3. Sufficient data and excellent generalization ability: The oil flow field and temperature field data under alternating loads of different frequencies are generated through CFD simulation, which makes up for the lack of field measured alternating load data, enhances the training dataset, and effectively improves the generalization ability of the model, making it suitable for alternating load conditions of different frequencies and amplitudes.

[0025] 4. Excellent real-time performance and strong engineering applicability: The model can be deployed on edge computing gateways to achieve local processing of real-time load data and millisecond-level prediction of hot spot temperature. It also outputs thermal aging coefficient and graded early warning signals, providing intuitive decision-making basis for on-site operation and maintenance personnel. It is suitable for practical engineering applications in new energy power plants.

[0026] 5. High interpretability, taking into account both mechanism and data: This invention is not a simple "black box" data-driven model, but integrates the physical mechanism of transformer heat conduction. The training process and prediction results of the model have clear physical meanings, which are easy for engineers to understand and apply, and also provide a clear direction for subsequent model optimization and improvement. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is an architecture diagram of the TCN-Attention hybrid model of the present invention; Figure 3 This invention provides a three-dimensional geometric model and temperature contour plot of a transformer obtained through CFD simulation. Figure 4 This is a comparison chart of the hotspot temperature prediction results of the present invention; Figure 5 This is a diagram showing the overall module architecture of the system of the present invention.

[0028] Explanation of reference numerals in the attached figures: Figure 2 In the middle: 1-Input layer, 2-TCN feature extraction layer, 3-Attention mechanism layer, 4-Output layer, 5-Physical constraint embedding module, 6-Joint loss function calculation module; Figure 5 In the middle: 10-Data construction module, 11-CFD simulation unit, 12-Data acquisition unit, 13-Data preprocessing unit, 20-Model design module, 21-TCN construction unit, 22-Attention mechanism construction unit, 30-Model training module, 31-Physical constraint embedding unit, 32-Training optimization unit, 40-Prediction output module, 50-Aging evaluation module, 60-Edge deployment module. Detailed Implementation

[0029] Example 1 like Figure 1-5 As shown, a method for dynamically predicting the hot spot temperature of a transformer under alternating load includes: S1. Construct a training dataset under alternating transformer load; S2. Design a Temporal Convolutional Network Attention Mechanism (TCN-Attention Hybrid Model); S3. Embed the transformer heat conduction equation as a physical constraint into the loss function of the TCN-Attention hybrid model to train a hotspot temperature prediction model; S4. Input the real-time alternating load data of the transformer into the trained hot spot temperature prediction model, and output the short-term predicted trajectory of the transformer winding hot spot temperature. S5. Calculate the thermal aging coefficient of transformer insulation paper based on the hot spot temperature prediction trajectory.

[0030] By constructing a composite dataset containing both simulation and measured data, and combining deep learning models with physical constraints, accurate perception of the thermal state of transformers under complex load conditions was achieved. The core logic of this method lies in using a TCN-Attention hybrid model to capture the long-sequence fluctuation characteristics generated by alternating loads, while embedding the underlying physical mechanism of transformer heat conduction into the training process as a loss function constraint. This design effectively overcomes the problems of lag and large inertial errors in traditional thermal circuit models under alternating loads, and also solves the defects of pure data-driven models, such as predictions that violate physical laws and diverge under data scarcity or extreme conditions. Through steps S1 to S5, not only can a high-precision temperature prediction trajectory be output, but also quantitative indicators of insulation aging can be further derived, providing closed-loop technical support from real-time monitoring to life assessment for transformer operation and maintenance in new energy scenarios.

[0031] In the preferred embodiment, step S1 involves constructing a training dataset under alternating transformer load, and the steps include: 1a. Establish a three-dimensional geometric model of the transformer winding, oil passage and insulation structure, set the alternating load waveforms of different frequencies as simulation boundary conditions, solve the oil flow field and temperature field distribution through CFD simulation, and extract the temperature time series data of the hot spot location of the winding. The training dataset contains transformer oil flow field and temperature field data under alternating loads at different frequencies generated by CFD simulation, as well as historical transformer operating data; 1b. Collect historical operating data of the transformer, which includes at least the load current sequence, ambient temperature sequence, top oil temperature sequence, and measured hot spot temperature sequence of the winding; 1c. The CFD simulation data and historical running data are preprocessed, fused, and divided into training set, validation set, and test set to obtain the training dataset.

[0032] The core logic of constructing the training dataset for transformer alternating load in step S1 lies in solving the problem of incomplete feature coverage by a single data source under extreme alternating load conditions through multi-source data fusion. Specifically, step 1a first establishes a three-dimensional geometric model of the transformer windings, oil channels, and insulation structure. Based on this, alternating load waveforms of different frequencies are set as boundary conditions for the simulation. Let the set alternating load current sequence be... Its corresponding alternating frequency is The time variable is According to the principles of electromagnetics and heat transfer, the heat loss generated by this alternating current is... With current sequence It is proportional to the square of the equation. By coupling and solving the continuity equation, momentum equation, and energy equation under the above boundary conditions using computational fluid dynamics simulation software, the velocity distribution of the oil flow field inside the transformer can be obtained. and temperature field distribution ,in , , This represents a three-dimensional spatial coordinate variable. Subsequently, the location of the highest winding temperature is extracted from the multi-dimensional calculation results, and the temperature sequence at this location over time is recorded as the temperature time-series data of the winding hotspot location. The beneficial effect of this step is that it can simulate extreme high-frequency and low-frequency alternating load conditions that are difficult to capture or have not yet occurred in actual transformer operation, and obtain the complete dynamic physical continuity of heat conduction and convection heat dissipation inside the transformer. This provides high-quality feature samples containing underlying thermophysical mechanisms for deep learning models, effectively avoiding the prediction divergence of purely data-driven models under unseen operating conditions, meeting the technical requirement of full disclosure, and providing underlying theoretical support for subsequent claims.

[0033] For step 1b, collecting historical transformer operating data is crucial for constructing a true physical baseline. The collected sequences specifically include load current sequences. Ambient temperature sequence Top oil temperature sequence and measured sequence of winding hot spot temperature In these variable definitions, This represents the actual load evolution level that the transformer experiences in a real power grid. It reflects the dynamic changes in the external natural heat dissipation environment. and This directly characterizes the actual thermal response state of the transformer under the dual influence of electrothermal coupling and environmental interference. This multi-dimensional operational data acquisition has significant beneficial effects. Historical operational data naturally includes implicit characteristics such as sensor measurement noise, real-world environmental climate interference, and material thermal aging after long-term transformer operation. Combining these field-measured data with the aforementioned simulation data as a training basis forces the prediction model to learn the real physical degradation laws and environmental coupling noise, ensuring that the final prediction model not only remains at the level of ideal simulation experiments but also accurately reflects the actual engineering site, greatly improving the prediction accuracy and robustness of the dynamic hotspot temperature prediction method in real power grid environments.

[0034] In step 1c, the data generated by the simulation and the historical operational data collected on-site are preprocessed and fused. The preprocessing mainly involves imputing missing values ​​in the original data sequence and eliminating dimensional differences between parameters, enabling data features from different sources and with different physical meanings to be compared, mapped, and used for loss calculation within the same numerical range. Let the total fused dataset be... This total dataset encompasses both ideal broadband time-series features extracted from simulations and noisy real-world operational features collected on-site. The total dataset will then be... The training set is divided according to a predetermined ratio. Validation set and test set In the model's execution logic, the training set... Used for backpropagation and iterative update of network weights within the model, validation set. Used to monitor model performance in real time during training to optimize hyperparameters and prevent network overfitting; test set This is then used to independently evaluate the final generalization prediction accuracy of the model after training is fully completed. The beneficial effect of this step is that, through data-level fusion technology, the dataset upon which the model training depends possesses both the true physical attributes of historical data and the comprehensive boundary coverage of simulation data for complex alternating operating conditions. This scientifically rigorous dataset partitioning and preprocessing mechanism strictly adheres to the normative paradigm of intelligent algorithms, ensuring the convergence of the model training process and the objectivity of the generalization ability evaluation.

[0035] The aforementioned complete implementation details elucidate the source of the training dataset, the mathematical transformation logic, and the final data format, enabling those skilled in the art to reproduce the technical solution unambiguously and completely eliminating the patent law risk of insufficient disclosure in the specification. Furthermore, this detailed, structured design at the underlying data foundation level provides rigorous and robust implementation support for the feature inputs of subsequent hybrid prediction models, ensuring that all claims are fully supported by the specification.

[0036] In the preferred embodiment, the data preprocessing in step 1c includes wavelet transform denoising, feature extraction, data normalization, and outlier removal performed sequentially. The extracted features include the high-frequency fluctuation component of the load current, the rate of change of current, and the temperature gradient features.

[0037] In the preferred scheme, step S2 involves designing the TCN-Attention hybrid model, which includes the following steps: 2a. Construct a TCN feature extraction layer, using stacked causal dilated convolution kernels to form a multi-level network structure, extract the time-dependent features of long alternating load sequences, and solve the gradient vanishing problem through residual connections; 2b. Connect an attention mechanism layer to the output of the TCN feature extraction layer to calculate the weight distribution of load features at different time steps and focus on the key feature points of the alternating load peaks and troughs. 2c. Construct the model output layer and map the output features of the attention mechanism layer to the predicted values ​​of transformer hot spot temperature.

[0038] The core logic of the data preprocessing step 1c lies in eliminating non-ideal factors mixed in the original data and extracting the core variables with the most physical representation significance. In the sequential steps, wavelet transform denoising first decomposes the original time-series signal containing background noise into different frequency scales. Let the original signal be x(t), and the wavelet basis function be ∑_{a,b}(t), where a is the scale factor and b is the translation factor. The continuous wavelet transform is expressed as W(a,b) = ∑_{a,b}(t) ... By applying thresholding to wavelet coefficients at multiple scales and then reconstructing them, white noise such as high-frequency electromagnetic interference can be effectively filtered out while preserving the edge features of signal abrupt changes. Subsequently, in the feature extraction stage, the high-frequency fluctuation component of the load current, the rate of change of current, and the temperature gradient features were extracted. The high-frequency fluctuation component of the current reflects the instantaneous load impact brought by the access of new energy sources; the rate of change of current, i.e., the first derivative of current with respect to time dI(t) dt, directly characterizes the speed of load abrupt changes; the temperature gradient reflects the spatial imbalance between heat accumulation and dissipation inside the transformer. Next, through data normalization, all extracted heterogeneous features are linearly mapped to the same dimension interval (usually between 0 and 1) to accelerate the gradient descent convergence process of the subsequent deep learning model and eliminate the weight bias caused by the large difference in numerical magnitude. Finally, outlier removal is performed, and absurd values ​​caused by occasional sensor failures are removed using statistical criteria. The combination of these sequential preprocessing steps has significant beneficial effects. It not only greatly purifies the sample space for model learning and ensures that the data input to the model are highly reliable and contain refined features with key evolutionary information, but also lays a solid data quality foundation for the model to accurately predict transient thermal responses under high-frequency alternating conditions, fully supporting the subsequent claims.

[0039] In step S2 of the preferred scheme, a hybrid model of temporal convolutional network attention mechanism (TCN-Attention) is designed. In step 2a, a temporal convolutional network (TCN) feature extraction layer is first constructed. The TCN layer abandons the inherent serial computation drawback of traditional recurrent neural networks and uses stacked causal dilated convolutional kernels to form a multi-level network structure. Causal convolution means that the output at the current time depends only on the input at the current and past time steps, strictly guaranteeing the physical causality of temporal prediction; while dilated convolution introduces holes between the convolutional kernel elements, so that the receptive field expands exponentially as the number of network layers increases. Let the th... The dilated convolution operation of the layer is ,in Given the input sequence, For size convolution kernel, As the expansion factor, The TCN uses sequence indexing. This structure allows for extremely efficient extraction of long-range temporal dependencies inherent in long sequences of alternating loads. Furthermore, to avoid gradient vanishing and degradation issues caused by network deepening, the TCN layer incorporates residual connections, enabling smooth backpropagation of gradient flows across layers. This design significantly enhances the model's ability to remember and extract features from complex, long-period fluctuations, providing a more macroscopic and profound historical perspective for predicting hotspot temperatures.

[0040] In step 2b, an attention mechanism layer is connected to the output of the TCN feature extraction layer. While the TCN layer extracts rich global temporal features, not all historical data points are equally important for predicting temperature at a future instant. The attention mechanism layer addresses this issue by calculating the weight distribution of the load features at different time steps. Its core computation process can be represented as first calculating the query vector... Key vector Sum value vector The formula for calculating the attention weight matrix is ​​usually as follows: ,in The dimension scaling factor for the key vector is used to prevent the gradient from vanishing due to excessively large dot product values. Through this calculation, the model can adaptively assign extremely high attention weights to critical time points in alternating load waveforms, such as sharp peaks and troughs, while suppressing the influence of stationary or redundant time steps. The beneficial effect of this mechanism is that it gives the model the ability to focus on core heating events, significantly improving the predictive sensitivity and accuracy at moments of temperature abrupt change under highly variable load conditions.

[0041] In step 2c, the model output layer is constructed. This layer takes the high-dimensional output feature sequence, which has been filtered and weighted by the attention mechanism layer, and performs nonlinear mapping through a fully connected network, ultimately reducing the dimensionality and integrating it into a specific predicted value for transformer hotspot temperature. The mapping formula for a fully connected layer is generally expressed as: ,in The output feature vector of the attention layer. This is the output layer weight matrix. This is the bias term. This step completes the final transformation from the abstract temporal feature space to the specific physical quantity (temperature value) space. Its beneficial effect is that it forms a complete end-to-end architecture closed loop from the input of the original temporal signal to the output of a specific physical target quantity. The entire TCN-Attention model is logically rigorous and progressively advanced, possessing both macroscopic long-range memory retrieval capabilities and microscopic key waveform focusing capabilities, providing a highly innovative and well-supported technical solution for solving the complex temperature prediction problem under alternating loads.

[0042] In the preferred embodiment, step S3 embeds the transformer heat conduction equation as a physical constraint into the loss function. The steps include: 5a. Based on the heat capacity-thermal resistance model of the transformer oil-paper insulation system, the heat conduction equation is determined as follows: ,in For equivalent heat capacity, For equivalent thermal resistance, For hot spot temperature, For ambient temperature, The loss source term varies with the square of the load current; 5b. Construct the joint loss function ,in The mean square error between the model's predicted values ​​and the measured values. This represents the sum of squared residuals of the heat conduction equation in a discrete grid. This is a balancing coefficient used to adjust the weights of data fitting and physical constraints. 5c. The backpropagation algorithm of deep reinforcement learning is used in combination with the Adam optimizer to train the model until the validation set loss converges.

[0043] Regarding the specific implementation process of embedding the transformer heat conduction equation as a physical constraint into the loss function in step S3, its core logic lies in breaking the black-box limitations of purely data-driven models and directly solidifying the thermodynamic conservation law into the boundary conditions for neural network optimization. In step 5a, based on the heat capacity and thermal resistance model of the transformer oil-paper insulation system, the specific mathematical expression of the heat conduction equation is established, which is expressed as: In this mathematical formula, The equivalent heat capacity represents the transformer windings and the surrounding insulating medium, reflecting the material's ability to absorb or release heat to change its own temperature. It represents the equivalent thermal resistance between the internal hot spot of the transformer and the external environment, and characterizes the degree of obstruction in the heat transfer process; This represents the hot spot temperature of the transformer winding to be solved; Represents a time variable; This represents the external ambient temperature of the transformer. This represents the loss source term that varies with the square of the load current, i.e., the dynamic heat input generated inside the transformer due to Joule heating. The beneficial effect of this design is that it provides a precise mathematical model of the complex electrothermal coupling physical processes inside the transformer through rigorous calculus equations. This gives subsequent models clear guidance from first principles of physics during learning, greatly limiting absurd predictions that might violate thermodynamic principles when facing abnormal load fluctuations. It provides solid theoretical support for improving the model's predictive reliability under extreme conditions, thus making the published content detailed and comprehensive.

[0044] In step 5b, to balance the fit between historical data and the underlying physical laws during model training, a joint loss function is constructed. The formula for this joint loss function is as follows: In this joint loss function formula, It represents the mean square error between the model's predicted values ​​and the actual measured values, and is mainly responsible for guiding the model network to approximate the true historical data distribution. This represents the sum of squared residuals generated by solving the heat conduction equation in a discrete-time grid, which is the accumulation of the unbalance on both sides of the equation after substituting the predicted temperature time series curve into the aforementioned physical equation. It is responsible for supervising that the prediction results of the model strictly follow the law of conservation of energy. The balancing coefficient adjusts the relative weight between data fitting error and physical constraint penalty during dynamic training. This joint loss function design has significant benefits, cleverly transforming the traditional unconstrained optimization problem into a constrained optimization problem with strong physical priors. This allows the model to maintain the smoothness and convergence of the predicted trajectory even when some sensor data is missing or subjected to strong electromagnetic interference, relying on the constraint of the physical loss term. This completely solves the core pain point of poor generalization ability in pure deep learning models and fully supports the specification requirement of integrating physical constraints into the algorithm's underlying structure.

[0045] In step 5c, the specific optimization strategy for model training is clarified: the backpropagation algorithm of deep reinforcement learning is used, combined with the Adam optimizer, to iteratively train the constructed hybrid model until the loss function on the validation set reaches a stable convergence state. The Adam optimizer can adaptively calculate the independent learning rate of each parameter, exhibiting extremely high convergence efficiency and optimization accuracy when dealing with non-stationary time-series data such as alternating transformer loads, which contain a large amount of noise and sparse gradients. The beneficial effect of this step is that by setting a scientifically rigorous gradient descent mechanism and stopping criteria, overfitting of the model is prevented, ensuring that the final deployed prediction model can find the globally optimal solution that satisfies both data and physical constraints in the multi-dimensional parameter space, thus guaranteeing the feasibility and reproducibility of the patented technology solution on actual computing platforms.

[0046] For the input of real-time alternating load data from the transformer and the prediction of short-term hot spot temperatures in the preferred scheme, a prediction time window spanning five minutes to seventy-two hours was set. This time scale not only covers the millisecond-level monitoring requirements for short-term thermal safety during sudden load surges but also takes into account the formulation cycle of day-ahead dispatching and medium- to long-term load transfer strategies for the power grid. The real-time alternating load data input to the model comprehensively covers real-time collected three-phase load current, ambient temperature, and top oil temperature. Before inputting these multi-source real-time streaming data into the model, a data preprocessing process completely consistent with that in the training set construction phase is strictly implemented. The beneficial effect of this front-end and back-end consistency processing mechanism is that it eliminates prediction distortion caused by data feature distribution shifts between the real-time online prediction and offline training phases. This allows the maturely trained hybrid model to seamlessly connect to the online monitoring system at the transformer site, outputting highly accurate temperature prediction trajectories in real time. This provides direct and reliable data decision-making basis for guiding power grid operation and maintenance personnel in load management and equipment life extension, and provides solid implementation support for the engineering implementation of the entire prediction method.

[0047] In the preferred scheme, the short time in S4 is 5min~72h. The real-time alternating load data of the transformer includes the real-time collected three-phase load current, ambient temperature and top oil temperature. Before inputting it into the model, it needs to be preprocessed in the same way as in S1.

[0048] Regarding the specific implementation process of setting the short-term forecast range in step S4 to five minutes to seventy-two hours, the core logic lies in constructing a multi-timescale thermal safety early warning mechanism. Setting the lower limit of the forecast window to five minutes accurately matches the real-time scheduling cycle of the power system's automatic generation control commands, providing millisecond- to minute-level instantaneous hotspot temperature monitoring for transformers to cope with sudden short-term overloads, preventing damage to winding insulation due to instantaneous heat accumulation. Setting the upper limit of the forecast window to seventy-two hours fully aligns with the time span of day-ahead and intraday power forecasts for new energy power plants, enabling the operation and maintenance system to combine the weather forecast for the next three days with the new energy power output plan curve to plan load transfer strategies or reserve backup cooling resources in advance. The beneficial effect of this multi-scale time window design is that it not only ensures the survivability of the underlying physical equipment of the power grid in the face of transient high current impacts, but also significantly improves the foresight and disaster prevention and mitigation efficiency of the top-level operation and scheduling, providing a comprehensive and ample decision-making buffer space for the full life-cycle health management of transformers under alternating loads.

[0049] In the acquisition of real-time alternating load data for transformers, it is explicitly required that all three-phase load current, ambient temperature, and top oil temperature be comprehensively covered in real time. To accurately represent this multi-dimensional time series acquisition process, let the current sampling time be... The input data vector can be represented as Its composition is In this matrix expression, Representative at The original multidimensional state feature vector collected by the front-end sensors of the time system; , , These represent the real-time alternating load currents that actually flow through phases A, B, and C on the high-voltage or low-voltage side of the transformer, respectively. The sum of the squares of the effective values ​​of these three currents directly determines the dynamic heat generation of the Joule losses in the transformer windings, which is the core internal cause of sudden temperature changes. The real-time temperature representing the natural environment in which the transformer is located constitutes the key thermodynamic boundary condition that determines the rate of heat dissipation and convection exchange from the surface of the equipment to the outside. This parameter represents the real-time measured temperature of the top insulating oil inside the transformer tank. As the temperature of the intermediate medium connecting the heat source and the heat dissipation end, it comprehensively reflects the overall thermal inertia reference state of the transformer under the coupling effect of the current load and environment. The beneficial effect of this characteristic input combination is that it comprehensively and without physical redundancy encompasses all the core internal and external factors driving changes in transformer hotspot temperature. This allows the prediction model to perform dynamic reasoning within a complete and closed thermodynamic constraint space, completely eliminating prediction curve distortion caused by the omission of key heating factors, and providing complete field data source support for high-precision nonlinear mapping of hotspot temperature.

[0050] Before feeding the aforementioned real-time data into the prediction network, data preprocessing identical to that in step S1 must be performed before inputting it into the model. Let the preprocessing feature mapping function be... The standardized feature tensor after processing is This conversion process can be rigorously represented as In this mapping formula, the preprocessing feature mapping function... The system rigorously and sequentially integrates one-dimensional sequence operations, including wavelet transform denoising, high-frequency current fluctuation and temperature gradient feature extraction, maximum-minimum linear normalization, and outlier removal based on statistical thresholds. This process must maintain complete consistency with the scaling parameters, scaling factors, and filtering basis functions used in the offline training phase. This mandatory alignment requirement is of paramount engineering necessity and offers significant benefits. By establishing strong consistency in feature distribution between the online inference and offline training ends, it effectively eliminates data distribution offsets caused by sensor temperature drift, differences in units of measurement, or occasional communication errors in industrial settings. This fundamentally ensures that the real-time inflow of heterogeneous online data sequences can be accurately projected onto the high-dimensional feature hyperplane of the trained and converged deep reinforcement learning model. This allows complex hybrid prediction models to maintain lossless prediction accuracy and generalization robustness even under rapidly changing alternating load conditions in engineering settings. The detailed input definitions and preprocessing alignment specifications described above enable those skilled in the art to clearly and unambiguously deploy offline algorithms as online real-time prediction modules. This eliminates the shortcomings of insufficient disclosure in the specification from the engineering details of algorithm implementation and provides solid and irrefutable implementation support for the validity of the system closed loop in the claims.

[0051] In the preferred scheme, the thermal aging coefficient is calculated in S5 using the six-degree rule or the Montsinger chemical kinetic equation, specifically as follows: Based on the predicted hot spot temperature trajectory, the aging rate of the insulating paper at different time steps is calculated, and the cumulative thermal aging coefficient is obtained by integrating over time. When the predicted hot spot temperature exceeds the preset temperature threshold, or the cumulative thermal aging coefficient exceeds the preset aging threshold, a graded early warning signal for transformer overheating or life loss is generated.

[0052] The core logic of step S5, which involves calculating the thermal aging coefficient and generating graded early warning signals, lies in scientifically transforming the predicted transformer physical temperature trajectory into a chemical kinetic quantitative index characterizing the insulation life loss of the equipment. In practice, the six degrees rule or the Montsinger chemical kinetic equation is used to calculate the aging rate of the insulating paper at different time steps. The formula for calculating the aging rate of the insulating paper is expressed as follows: In this calculation formula, This represents the thermal aging rate of the insulating paper; This represents the transformer winding hot spot temperature output by the prediction model; This represents the standard reference life temperature for the insulating paper. In practical engineering applications, the standard reference life temperature is typically set at 98 degrees Celsius. According to this chemical kinetic equation, for every six degrees Celsius increase in the predicted hot spot temperature, the thermal aging rate of the insulating paper will double, meaning the rate of life loss of the insulating material will double. This detailed design directly maps abstract temperature time-series values ​​to the degradation rate at the physicochemical level of the material, providing accurate and universally accepted theoretical support for subsequent life loss assessment.

[0053] Based on the aging rates of the insulating paper at different time steps obtained above, the cumulative thermal aging coefficient of the transformer within the predicted period can be obtained by integrating or discretizing and accumulating over the time dimension. The formula for calculating the cumulative thermal aging coefficient is expressed as follows: In this calculation formula, It represents the cumulative thermal aging coefficient, used to quantify the total relative loss of transformer insulation life over a continuous period of time; Represents the index number of the discrete time step in the time series; This represents the total number of time steps contained in the temperature prediction trajectory. Representing the The instantaneous aging rate of the insulating paper corresponding to each time step; This represents the time interval between two adjacent time steps. Through this discrete integral accumulation calculation method, the system can accurately capture the minute and irreversible thermal damage to the insulating paper caused by each temperature peak caused by wind and solar fluctuations under alternating load conditions, thereby achieving fine tracking and precise quantitative calculation of the cumulative consumption process of transformer insulation life.

[0054] After obtaining the predicted future hotspot temperature and cumulative thermal aging coefficient, the system compares these two state parameters with pre-set safety thresholds to trigger and generate tiered early warning signals. The preset temperature threshold is set as follows: Set the preset aging threshold as When the future hotspot temperature forecast value Exceeding the preset temperature threshold When the system determines that the transformer faces a serious risk of transient thermal runaway or winding insulation burnout, it immediately generates a Level 1 warning signal for transient overheating. When the cumulative thermal aging coefficient... Exceeding the preset aging threshold When the system determines that the transformer has undergone severe irreversible degradation of its internal insulation paper due to heat accumulation under long-term complex alternating load conditions, it generates a level-two warning signal for lifespan loss. If both of the above conditions exceeding the threshold are met simultaneously, the system generates the highest-level level-three warning signal, requiring immediate load disconnection or emergency shutdown. This multi-dimensional, hierarchical warning mechanism based on dual-parameter comparison helps power grid maintenance personnel accurately distinguish between sudden extreme overheating hazards and long-term cumulative material fatigue hazards.

[0055] The aforementioned chemical kinetic calculations and the implementation of the multi-level early warning mechanism have yielded significant beneficial effects. This scheme completely breaks through the lag limitation of traditional transformer relay protection systems, which only rely on the current threshold exceeding the limit for passive disconnection response, achieving a major technological leap towards proactive and predictive condition maintenance. By calculating the relative insulation life loss in future periods in advance and conducting dual-track threshold analysis, this mechanism provides the power dispatch center with sufficient safety response time, enabling the control system to issue load transfer commands or automatically activate auxiliary forced cooling and other proactive intervention measures before irreversible thermal damage actually occurs. This not only greatly reduces the sudden shutdown failure rate of new energy transformers under extreme fluctuation conditions but also enables the use of scientific data to support and guide the dynamic capacity expansion operation and accurate decommissioning assessment of transformers. This rigorous data deduction and physical rule mapping comprehensively ensures the safe and stable operation of core transformer equipment, providing a detailed, complete, and engineering-feasible technical path for the implementation of the aging calculation and graded early warning methods in the claims, effectively preventing the problem of insufficient disclosure due to abstract steps, and establishing a solid specification support system for the relevant claims.

[0056] In the preferred embodiment, the method is deployed in the edge computing gateway of the new energy power plant, and the real-time dynamic prediction of transformer hot spot temperature is achieved through edge computing; New energy power plants are wind power, photovoltaic power, or integrated wind-solar-storage power plants.

[0057] This application describes a specific implementation scheme for deploying the method of this application in an edge computing gateway of a new energy power plant to achieve real-time dynamic prediction of transformer hotspot temperature through edge computing. The core logic lies in offloading the computing power of the deep reinforcement learning prediction algorithm from the traditional centralized cloud to the edge of the device network closer to the data source. In specific engineering implementation, the converged prediction model trained on a high-performance server undergoes lightweight processing such as parameter pruning and quantization compression, and is then embedded in the edge computing gateway chip installed at the transformer site. The edge computing gateway directly receives real-time operating data collected by front-end sensors in parallel via a high-speed industrial bus. To quantify the real-time advantage of this deployment architecture, a time delay calculation model for the real-time response of edge computing is constructed as follows: In this time delay calculation formula, This represents the total latency of the system from acquiring underlying data to outputting prediction results and generating strategies. This represents the time consumed by the front-end sensor to collect three-phase load current and various temperature data and complete analog-to-digital conversion. The time consumption of a lightweight hotspot temperature prediction model to perform matrix multiplication and forward inference computations in the local processor of an edge computing gateway; This represents the communication time consumed when the predicted results and control commands are sent to the cooling actuator via the local area network.

[0058] This edge computing deployment architecture offers significant advantages. By eliminating the extremely time-consuming communication steps of traditional cloud-based prediction architectures, such as packaging massive amounts of high-frequency data, uploading over long distances over wide area networks, and queuing and parsing in the cloud, the total system latency is significantly reduced. The accuracy can be strictly controlled within extremely short milliseconds, thus truly achieving high-frequency, real-time dynamic prediction of hotspot temperatures. Simultaneously, the edge computing architecture significantly saves communication bandwidth resources for external transmission from renewable energy power plants, reducing reliance on the stability of the external network environment. Even in severe conditions where extreme weather causes the main communication network between the power plant and the dispatch center to be interrupted, the prediction system deployed at the edge gateway can still independently perform hotspot temperature inference based entirely on locally collected data, and autonomously trigger local overheat protection and cooling control actions in a closed-loop manner, ensuring the absolute thermodynamic safety of core transformer equipment during network outages.

[0059] For the specific application scenarios of wind power plants, photovoltaic power plants, or integrated wind-solar-storage power plants, as explicitly defined in this application, the underlying technical logic lies in accurately matching the unique and severe electrothermal conditions of these plants. The generator output of a wind power plant is directly constrained by drastic changes in wind speed; the inverter output power of a photovoltaic power plant is highly dependent on the real-time alternation of cloud cover and solar irradiance intensity; and integrated wind-solar-storage power plants involve a more complex process of frequent switching between energy storage charging and discharging. The inherent randomness, intermittency, and high volatility of these meteorological conditions and operating modes directly lead to the grid-connected main transformer enduring highly destructive, large-amplitude, high-frequency sawtooth alternating load impacts over long periods. Traditional transformer hotspot prediction models under stable operating conditions are prone to feature omissions and prediction divergences in such scenarios.

[0060] The application of the deeply customized dynamic prediction method in this application to the aforementioned new energy power plants yields the following benefits: it fully leverages the powerful ability of hybrid deep learning models to capture long-sequence extreme alternating load characteristics. The system can accurately identify and predict the transient heat accumulation process caused by sudden high-current surges triggered by strong winds or brief periods of intense sunlight. This effectively prevents accelerated aging or even insulation breakdown accidents in the grid-connected transformer insulation system under frequent thermal expansion and contraction and localized extreme high temperatures, significantly extending the actual service life of expensive main equipment. This highly targeted technical deployment completely addresses the safety shortcoming of lagging thermal status monitoring of power equipment during large-scale grid connection of new energy, greatly improving the stability and economic benefits of the underlying physical nodes of the power grid during the full consumption of clean energy. It provides solid, detailed, and irrefutable specification and implementation examples to support the large-scale industrial application of this invention patent in the field of new energy.

[0061] A dynamic prediction system for transformer hotspot temperature under alternating load based on deep reinforcement learning includes: The data construction module is used to build a training dataset for transformers under alternating loads, integrate CFD simulation data and historical transformer operating data, and complete data preprocessing. The model design module is used to design the TCN-Attention hybrid model and build a network architecture that includes a TCN feature extraction layer, an attention mechanism layer, and an output layer. The model training module is used to embed the transformer heat conduction equation as a physical constraint into the model loss function, and to train the hot spot temperature prediction model using a deep reinforcement learning algorithm. The prediction output module is used to receive real-time alternating load data of the transformer, input it into the trained prediction model, and output the short-term predicted trajectory of the hot spot temperature. The aging assessment module is used to calculate the thermal aging coefficient based on the hot spot temperature prediction trajectory and generate graded early warning signals according to preset thresholds.

[0062] In the preferred embodiment, the CFD simulation unit is used to establish a three-dimensional geometric model of the transformer, set boundary conditions for alternating loads at different frequencies, and simulate and generate oil flow field and temperature field data. The data acquisition unit is used to collect historical operating data and real-time alternating load data of the transformer through sensors; The data preprocessing unit is used to perform wavelet transform denoising, feature extraction, normalization, and outlier removal on simulation data and measured data; The model design module includes a TCN building unit and an attention mechanism building unit; The TCN building block is used to build a multi-level TCN network with causal dilated convolution, and to set residual connections and dilation factors. The attention mechanism building unit is used to design the weight calculation module and generate attention weight matrices for load features at different time steps; The model training module includes a physical constraint embedding unit and a training optimization unit; The physical constraint embedding unit is used to determine the heat conduction equation based on the transformer thermal capacity-thermal resistance model and construct a joint loss function that includes data fitting terms and physical residual terms. The training and optimization unit is used to train the model using the deep reinforcement learning backpropagation algorithm combined with the Adam optimizer, thereby updating and optimizing the network weights.

[0063] The prediction output module adopts a sequence-to-sequence encoding-decoding structure. The encoding end extracts the temporal features of the real-time load data, and the decoding end outputs the predicted hotspot temperatures at different time steps in the future, forming a continuous temperature prediction trajectory.

[0064] This paper addresses a dynamic prediction system for transformer hotspot temperature under alternating loads based on deep reinforcement learning. The system employs a modular architecture to achieve a complete prediction closed loop, from multi-source data fusion to insulation life assessment. The data construction module serves as the fundamental data source for the entire system, deeply integrating a computational fluid dynamics simulation unit, a data acquisition unit, and a data preprocessing unit. The computational fluid dynamics simulation unit establishes a detailed three-dimensional geometric model of the transformer windings, oil channels, and insulation structure, sets alternating load waveforms of different frequencies as simulation boundary conditions, and then generates global oil flow and temperature field data within the transformer through fluid dynamics solutions. The data acquisition unit acquires historical long-cycle operating data and current real-time alternating load data from various sensor nodes deployed on the transformer body. Subsequently, the data preprocessing unit performs wavelet transform denoising, feature extraction, normalization, and outlier removal operations on the simulation and measured data. This collaborative construction method using multi-source heterogeneous data has significant benefits. It not only compensates for the severe lack of field-measured samples under extreme alternating conditions through simulation but also completely eliminates environmental white noise and dimensional differences during sensor acquisition using preprocessing techniques. This provides a high-quality feature sample space that is distortion-free and covers all operating conditions for subsequent deep learning networks. It completely eliminates the problem of insufficient data disclosure caused by the single sample in model training from the underlying data level, and lays the data foundation for the stable operation of the entire system.

[0065] The model design module focuses on designing a hybrid model that integrates temporal convolutional networks and attention mechanisms. It consists of temporal convolutional network building units and attention mechanism building units. The temporal convolutional network building units are used to construct a multi-level network architecture with causal dilated convolutions, internally incorporating residual connections and dilation factors. The operation of causal dilated convolutions can be precisely represented as... In this formula, Represents the dilated convolution at the current sequence index. Output feature value at; The representative size is The first one in a one-dimensional convolution kernel Each weighted element; Represents the input alternating load time-series data sequence; The expansion factor controls the span distance of convolutional kernel elements in the time dimension. This structure allows the network to exponentially expand its receptive field without revealing any future time-step information, thereby deeply extracting hidden temporal dependencies in long sequences of alternating loads. Residual connections effectively overcome the vanishing gradient problem that is prone to occur during deep network training. The attention mechanism building unit is used to design the weight calculation module, generating attention weight matrices for load features at different time steps. The feature weighting formula for the attention mechanism can be expressed as: In this calculation formula, The query feature matrix contains the hidden state information of the system at the current time step. The key feature matrix represents the hidden state information of all historical time steps; The representative feature matrix contains historical features that need to be extracted by the network using weighted methods. The dimension value of the key feature matrix is ​​used to scale the dot product score to prevent computational overflow and gradient vanishing; softmax represents the normalized exponential activation function, responsible for transforming the scaled dot product score into a probability weight distribution that sums to one. The beneficial effect of this module design is that it endows the model with the ability to focus on core thermal events, enabling it to adaptively assign the highest weights to the peaks and troughs in the alternating load waveform that cause dramatic temperature changes. This greatly improves the model's feature mapping sensitivity and accuracy under drastically fluctuating conditions, providing solid specification and implementation support for the network topology structure claimed in the claims.

[0066] The model training module includes a physical constraint embedding unit and a training optimization unit. The physical constraint embedding unit determines the underlying heat conduction equation based on the heat capacity and thermal resistance model of the transformer oil-paper insulation system, and constructs a joint loss function containing data fitting terms and physical residual terms. The specific mathematical expression of the heat conduction equation is as follows: In this differential equation, It represents the equivalent heat capacity of the transformer winding and its surrounding insulating medium; This represents the equivalent thermal resistance between the internal hot spot of the transformer and the external natural environment. This represents the hot spot temperature of the transformer winding to be solved; Represents a continuous time variable; This represents the external ambient temperature of the transformer. This represents the loss source term that dynamically varies with the square of the alternating load current, i.e., the internal Joule heat generation. The joint loss function constructed based on this physical equation is expressed as: In this joint loss function formula, Loss represents the total loss value that needs to be minimized during model training; It represents the mean square error between the model's predicted temperature value and the actual measured temperature value, and is responsible for guiding the model to accurately fit the historical data distribution. This represents the sum of squared discrete residuals generated after substituting the predicted temperature curve into the heat conduction equation, and is responsible for ensuring that the prediction results strictly adhere to the law of conservation of energy. The adaptive balancing coefficient is used to dynamically adjust the weight ratio of data empirical fitting error and physical prior constraint penalty in the total loss. The training optimization unit uses a deep reinforcement learning backpropagation algorithm combined with an adaptive moment estimation optimizer to train the model, achieving rapid updating and smooth optimization of network weights. The beneficial effects of this training mechanism are extremely prominent. It breaks the black-box limitation of purely data-driven models, directly solidifying the law of thermodynamic conservation into a hard boundary condition for neural network optimization. This forces the model to find a globally optimal solution in a complex high-dimensional solution space that conforms to both statistical data and strictly adheres to the first principles of underlying physics. It completely solves the prediction divergence and logical absurdity problems that traditional algorithms are prone to when facing unseen extreme alternating loads.

[0067] The prediction output module and the aging assessment module together constitute the system's active defense and decision-making execution end. The prediction output module adopts a sequence-to-sequence encoding and decoding two-layer structure. In this structure, the encoding end is responsible for extracting the deep temporal features of the real-time input load data and compressing it into a high-density context semantic vector. The decoding end then continuously outputs the predicted hotspot temperature values ​​for different future time steps based on this context vector, thus seamlessly splicing them into a complete and continuous future temperature prediction trajectory. The aging assessment module is used to accurately calculate the thermal aging coefficient of the insulating paper based on the hotspot temperature prediction trajectory generated by the front-end module, and strictly generates graded early warning signals according to the pre-set safety red line threshold. The beneficial effect of this decision architecture design is that it transforms the hidden and complex electromagnetic and thermodynamic coupling black box calculations at the front end into an intuitive and visible future temperature evolution trend in real time, and then directly quantifies it into the relative lifespan deterioration rate of the transformer's core insulation material. When the predicted temperature trajectory or cumulative aging coefficient touches the preset red line, the system can immediately generate multi-dimensional graded early warning instructions. This completely changes the severe lag in traditional relay protection systems, which can only passively disconnect when the current temperature exceeds the limit. It provides the power grid dispatch and operation center with invaluable time for safety intervention and load transfer. The detailed module function division, internal algorithm logic analysis, and specific physical formula mapping described above enable those skilled in the art to clearly and accurately reproduce this hotspot temperature dynamic prediction system. This completely eliminates the insufficient disclosure caused by the abstract and vague descriptions of various functional modules, providing indisputable and solid support for the system structure definition features in the claims.

[0068] Example 2 Further explanation in conjunction with Example 1, such as Figure 1-5 The structure shown illustrates a method for dynamic prediction of transformer hotspot temperature under alternating load based on deep reinforcement learning. This embodiment combines Figure 1 The prediction method of the present invention is described in detail below, with specific steps as follows: 1. Construct the training dataset 1.1 CFD Simulation Data Generation: A three-dimensional geometric model of a 10kV / 400kVA wind-solar-storage integrated power station transformer was established using Fluent software. The model includes the high-voltage winding, low-voltage winding, oil channels, insulating paperboard, and tank structure. Unstructured meshing was used, with mesh refinement applied to the winding hotspot areas. Sawtooth alternating load waveforms at three typical frequencies (5Hz, 10Hz, and 15Hz) were set as boundary conditions to simulate the spatiotemporal distribution of the oil flow field and temperature field inside the transformer. Temperature time-series data at the winding hotspot location (upper 1 / 3 of the winding) were extracted, generating 1000 sets of "load-temperature" mapping data for each frequency.

[0069] 1.2 Historical Operation Data Acquisition: Historical operation data of the transformer for two consecutive years was collected, including timestamps, three-phase load current on the high-voltage side, ambient temperature, top oil temperature, and winding hot spot temperature (measured by fiber optic grating sensors), totaling 100,000 valid data points.

[0070] 1.3 Data Preprocessing and Fusion: CFD simulation data and historical operating data are preprocessed in sequence, including wavelet transform denoising (removing sensor noise and environmental interference), feature extraction (extracting high-frequency fluctuation components of load current, current change rate, and temperature gradient), data normalization (normalizing all features to the [0,1] interval), and outlier removal (using the 3σ principle to remove outlier data). The preprocessed simulation data and measured data are divided into training set, validation set, and test set in a 7:2:1 ratio to construct the training dataset.

[0071] 2. Design a TCN-Attention hybrid model This embodiment combines Figure 2 The design model is an end-to-end time-series prediction network. The input layer receives a time-series feature vector of length 96 (corresponding to 1-minute data points for 4 hours), and the output layer outputs the predicted hotspot temperature values ​​for the next hour. The specific structure is as follows: 2.1 TCN Feature Extraction Layer: Causal dilated convolution is used with a kernel size of 3 and dilation factors of 1, 2, 4 and 8 respectively. Four layers of TCN network are stacked, with 64 convolution kernels in each layer. The gradient vanishing problem is solved by residual connections to extract the time-dependent features of long alternating load sequences.

[0072] 2.2 Attention Mechanism Layer: An additive attention mechanism is adopted to map the feature vector output by TCN to query vector Q, key vector K and value vector V, calculate the attention weight matrix, and assign higher weights to key time steps such as load peaks and troughs to focus on the key features of alternating load.

[0073] 2.3 Output layer: A fully connected layer is used to map the output features of the attention mechanism layer to the hotspot temperature prediction values ​​for the next 60 time steps (1 minute / step), forming a continuous temperature prediction trajectory.

[0074] 3. Embed physical constraints and train the model 3.1 Determining the heat conduction equation: Based on the heat capacity-thermal resistance model of the transformer oil-paper insulation system, the heat conduction equation is determined as follows: ,in For equivalent heat capacity, For equivalent thermal resistance, For hot spot temperature, For ambient temperature, For load current, For winding copper resistance, loss source terms It varies with the square of the load current.

[0075] 3.2 Constructing the Joint Loss Function: Designing the Joint Loss Function ; in For data fitting terms, For actual measured temperature, Predict temperature for the model; is the physical residual of the heat conduction equation, and 0.3 is the equilibrium coefficient. .

[0076] 3.3 Model Training: The backpropagation algorithm of deep reinforcement learning is used in combination with the Adam optimizer (learning rate set to 0.001, batch size of 32) to train the model. During the training process, the physical residual is calculated by automatic differentiation. The training effect is monitored by the validation set loss. When the validation set loss no longer decreases for 10 consecutive epochs, the training is stopped, and the hotspot temperature prediction model is obtained after training.

[0077] 4. Dynamic prediction of hotspot temperature The real-time alternating load data of the transformer (three-phase load current, ambient temperature, top oil temperature) is input into the model. Before input, the same preprocessing as in step 1.3 is performed. The model outputs the winding hot spot temperature prediction trajectory for the next hour through a sequence-to-sequence encoding-decoding structure. The prediction process is completed in the edge computing gateway. The single-step prediction time is ≤50ms, achieving millisecond-level real-time prediction.

[0078] 5. Calculation and early warning of thermal aging coefficient Based on the predicted hotspot temperature trajectory, the thermal aging coefficient of the insulating paper is calculated using the six-degree rule: when the hotspot temperature is 98℃, the lifespan of the insulating paper is 20 years; for every 6℃ increase in temperature, the lifespan is halved. The calculation formula is as follows: ,in For aging rate, The hotspot temperature is used as the reference temperature. The cumulative thermal aging coefficient is obtained by integrating the aging rate over time. The temperature threshold is set at 110℃, and the aging coefficient threshold is 0.01 / day. When the predicted hotspot temperature is ≥110℃, a Level 1 overheat warning is triggered; when the cumulative thermal aging coefficient is ≥0.01 / day, a Level 2 lifespan loss warning is triggered; when both conditions are met simultaneously, a Level 3 emergency warning is triggered. The warning signal is pushed to maintenance personnel via sound and light and background SMS.

[0079] Example 3 Optimization design of boundary conditions in CFD simulation This embodiment optimizes the CFD simulation steps in Embodiment 1, and is applicable to alternating load conditions with different wind and solar fluctuation characteristics. Specifically: 1. Based on actual wind speed monitoring data of a wind farm, the probability distribution characteristics of wind speed fluctuations are extracted and transformed into the fluctuation characteristics of load current. The composite alternating load waveform of "sine wave + random noise" is set as the boundary condition for CFD simulation, which is closer to the actual load characteristics of wind and solar fluctuations.

[0080] 2. During the simulation, the variable physical properties of transformer oil are considered, namely the dynamic viscosity and thermal conductivity of transformer oil as a function of temperature. Variable physical property parameters are used to solve the oil flow field and temperature field to improve the realism of CFD simulation data.

[0081] 3. Extract hotspot temperature data under different load amplitudes (50%~150% of rated load) and different fluctuation frequencies (1Hz~20Hz) from the simulation results to generate a total of 5,000 sets of simulation data. After merging with the measured data, the richness of the training dataset is further improved, making the model applicable to a wider range of alternating load conditions.

[0082] Example 4 Lightweight design of TCN-Attention model This embodiment presents a lightweight design for the TCN-Attention model in Embodiment 1, making it suitable for edge computing gateways with limited hardware resources. Specifically: 1. Prune the TCN feature extraction layer by reducing the number of convolutional kernels from 64 to 32, removing one dilated convolutional layer, and retaining three dilated convolutional layers (dilation factors 1, 2, and 4) to reduce the number of model parameters.

[0083] 2. Depthwise separable convolution is used instead of traditional convolution, separating spatial convolution and channel convolution, further reducing the computational cost of the model.

[0084] 3. The attention mechanism layer is simplified by replacing the additive attention mechanism with a dot product attention mechanism to reduce the complexity of weight calculation.

[0085] The lightweight model reduces the number of parameters by 60%, the amount of computation by 55%, and the single-step prediction time is ≤30ms. It can achieve more efficient real-time prediction in edge computing gateways, and the prediction accuracy only decreases by 0.5℃, which meets the accuracy requirements of actual engineering.

[0086] Example 5 A Dynamic Prediction System for Transformer Hotspot Temperature under Alternating Load Based on Deep Reinforcement Learning This embodiment combines Figure 5 The prediction system of the present invention is described in detail below. The system includes a data construction module 10, a model design module 20, a model training module 30, a prediction output module 40, an aging assessment module 50, and an edge deployment module 60. The functions of each module are as follows: 1. Data Construction Module 10: Includes a CFD simulation unit 11, a data acquisition unit 12, and a data preprocessing unit 13. The CFD simulation unit 11 is used to establish a three-dimensional geometric model of the transformer, set the boundary conditions of alternating loads with different frequencies and waveforms, and simulate and generate oil flow field and temperature field data; the data acquisition unit 12 collects historical operating data and real-time alternating load data of the transformer through fiber optic grating sensors, current transformers, and temperature sensors; the data preprocessing unit 13 is used to perform wavelet transform denoising, feature extraction, normalization, and outlier removal on the simulation data and measured data, and complete data fusion and dataset partitioning.

[0087] 2. Model Design Module 20: Includes TCN building unit 21 and attention mechanism building unit 22. TCN building unit 21 is used to build a multi-level TCN network with causal dilated convolution, and set residual connections, convolution kernel size and dilation factor; attention mechanism building unit 22 is used to design the attention weight calculation module, generate attention weight matrix of features loaded at different time steps, and focus on key feature points.

[0088] 3. Model Training Module 30: Includes a physical constraint embedding unit 31 and a training optimization unit 32. The physical constraint embedding unit 31 determines the heat conduction equation based on the transformer thermal capacity-thermal resistance model and constructs a joint loss function containing data fitting terms and physical residual terms. The training optimization unit 32 uses the backpropagation algorithm of deep reinforcement learning, combined with the Adam optimizer, to train the model, realize the updating and optimization of network weights, until the validation set loss converges.

[0089] 4. Prediction output module 40: It adopts a sequence-to-sequence encoding-decoding structure, receives real-time alternating load preprocessing data transmitted by data construction module 10, inputs it into the trained prediction model, extracts time-series features at the encoding end, and outputs the predicted hot spot temperature values ​​for the next short time at the decoding end, forming a continuous temperature prediction trajectory.

[0090] 5. Aging Assessment Module 50: Based on the temperature prediction trajectory of the prediction output module 40, the thermal aging coefficient is calculated using the six degrees rule or the Montsinger chemical kinetic equation. The calculation results are compared with the preset temperature threshold and aging coefficient threshold to generate first-level, second-level, and third-level graded early warning signals, which are then pushed to the operation and maintenance terminal.

[0091] 6. Edge Deployment Module 60: The trained prediction model is converted into a lightweight form and deployed on the edge computing gateway of the new energy power station to realize local processing of real-time load data and millisecond-level prediction of hot spot temperature. It also supports online updates and iterative optimization of the model. When new operating data is collected, incremental training of the model can be realized.

[0092] The various modules of this system are interconnected through data interfaces. Data transmission adopts the industrial Ethernet protocol to ensure the real-time performance and reliability of data transmission. The system can be connected to the power plant's SCADA system to realize remote monitoring and centralized management of the transformer's thermal status.

[0093] Attached Figure Description (Detailed Explanation) Figure 1 The flowchart of the present invention clearly shows the entire process from dataset construction, model design, physical constraint embedding, model training, to hotspot temperature prediction, thermal aging assessment and early warning. Each step is connected in sequence to form a complete prediction method system.

[0094] Figure 2 The diagram shows the architecture of the TCN-Attention hybrid model of this invention, illustrating the four-layer structure of the model: input layer 1 receives temporal feature vectors, TCN feature extraction layer 2 extracts long sequence features through causal dilated convolution, attention mechanism layer 3 calculates feature weights, and output layer 4 outputs temperature prediction values; it also shows the physical constraint embedding module 5 and the joint loss function calculation module 6, realizing the fusion of physical constraints and model training.

[0095] Figure 3 The image shows the three-dimensional geometric model and temperature cloud map of the transformer used in the CFD simulation of this invention. The left side shows the three-dimensional geometric model and mesh generation effect of the transformer winding and oil passage, while the right side shows the temperature cloud map inside the transformer under alternating load, clearly showing the temperature distribution at the hot spot location of the winding, providing a basis for the extraction of simulation data.

[0096] Figure 4 This is a comparison chart of the hotspot temperature prediction results of the present invention. The horizontal axis represents time (minutes), and the vertical axis represents the hotspot temperature (°C). The solid line represents the measured value, the dashed line represents the predicted value of the present invention, the dotted line represents the predicted value of the traditional thermal circuit model, and the dotted-dash line represents the predicted value of the pure LSTM model. As can be seen from the figure, the predicted value of the present invention is highly consistent with the measured value with small error and can accurately track the temperature change of alternating load. In contrast, the traditional thermal circuit model has obvious lag, and the pure LSTM model shows prediction deviation when the load increases sharply.

[0097] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for dynamically predicting the hot spot temperature of a transformer under alternating load, characterized in that: the method include: S1. Construct a training dataset under alternating transformer load; S2. Design a Temporal Convolutional Network Attention Mechanism (TCN-Attention Hybrid Model); S3. Embed the transformer heat conduction equation as a physical constraint into the loss function of the TCN-Attention hybrid model to train a hotspot temperature prediction model; S4. Input the real-time alternating load data of the transformer into the trained hot spot temperature prediction model, and output the short-term predicted trajectory of the transformer winding hot spot temperature. S5. Calculate the thermal aging coefficient of transformer insulation paper based on the hot spot temperature prediction trajectory.

2. The method for dynamically predicting transformer hot spot temperature under alternating load according to claim 1, characterized in that: Step S1 involves constructing a training dataset under alternating transformer load, including the following steps: 1a. Establish a three-dimensional geometric model of the transformer winding, oil passage and insulation structure, set the alternating load waveforms of different frequencies as simulation boundary conditions, solve the oil flow field and temperature field distribution through CFD simulation, and extract the temperature time series data of the hot spot location of the winding. The training dataset contains transformer oil flow field and temperature field data under alternating loads at different frequencies generated by CFD simulation, as well as historical transformer operating data; 1b. Collect historical operating data of the transformer, which includes at least the load current sequence, ambient temperature sequence, top oil temperature sequence, and measured hot spot temperature sequence of the winding; 1c. The CFD simulation data and historical running data are preprocessed, fused, and divided into training set, validation set, and test set to obtain the training dataset.

3. The method for dynamically predicting transformer hot spot temperature under alternating load according to claim 2, characterized in that: The data preprocessing described in step 1c includes wavelet transform denoising, feature extraction, data normalization, and outlier removal performed sequentially. The extracted features include the high-frequency fluctuation component of the load current, the rate of change of current, and temperature gradient features.

4. The method for dynamically predicting transformer hot spot temperature under alternating load according to claim 1, characterized in that: Step S2 involves designing the TCN-Attention hybrid model, which includes the following steps: 2a. Construct a TCN feature extraction layer, using stacked causal dilated convolution kernels to form a multi-level network structure, extract the time-dependent features of long alternating load sequences, and solve the gradient vanishing problem through residual connections; 2b. Connect an attention mechanism layer to the output of the TCN feature extraction layer to calculate the weight distribution of load features at different time steps and focus on the key feature points of the alternating load peaks and troughs. 2c. Construct the model output layer and map the output features of the attention mechanism layer to the predicted values ​​of transformer hot spot temperature.

5. The method for dynamically predicting transformer hot spot temperature under alternating load according to claim 1, characterized in that: Step S3 embeds the transformer heat conduction equation as a physical constraint into the loss function. The steps include: 5a. Based on the heat capacity-thermal resistance model of the transformer oil-paper insulation system, the heat conduction equation is determined as follows: ,in For equivalent heat capacity, For equivalent thermal resistance, For hot spot temperature, For ambient temperature, The loss source term varies with the square of the load current; 5b. Construct the joint loss function ,in The mean square error between the model's predicted values ​​and the measured values. This represents the sum of squared residuals of the heat conduction equation in a discrete grid. This is a balancing coefficient used to adjust the weights of data fitting and physical constraints. 5c. The backpropagation algorithm of deep reinforcement learning is used in combination with the Adam optimizer to train the model until the validation set loss converges.

6. The method for dynamically predicting the hot spot temperature of a transformer under alternating load according to claim 1, characterized in that: S4 The short-term timeframe is 5 minutes to 72 hours. The real-time alternating load data of the transformer includes real-time collected three-phase load current, ambient temperature, and top oil temperature. Before inputting the data into the model, it needs to undergo the same preprocessing as in S1.

7. The method for dynamically predicting the hot spot temperature of a transformer under alternating load according to claim 1, characterized in that: the thermal aging coefficient is calculated in S5 using the six-degree rule or the Montsinger chemical kinetic equation, specifically as follows: Based on the predicted hot spot temperature trajectory, the aging rate of the insulating paper at different time steps is calculated, and the cumulative thermal aging coefficient is obtained by integrating over time. When the predicted hot spot temperature exceeds the preset temperature threshold, or the cumulative thermal aging coefficient exceeds the preset aging threshold, a graded early warning signal for transformer overheating or life loss is generated.

8. A method for dynamically predicting transformer hot spot temperature under alternating load according to any one of claims 1-7, characterized in that: the method Deployed in the edge computing gateway of new energy power plants, it enables real-time dynamic prediction of transformer hot spot temperature through edge computing; New energy power plants are wind power, photovoltaic power, or integrated wind-solar-storage power plants.

9. A dynamic prediction system for transformer hotspot temperature under alternating load based on deep reinforcement learning, characterized in that: include: The data construction module is used to build a training dataset for transformers under alternating loads, integrate CFD simulation data and historical transformer operating data, and complete data preprocessing. The model design module is used to design the TCN-Attention hybrid model and build a network architecture that includes a TCN feature extraction layer, an attention mechanism layer, and an output layer. The model training module is used to embed the transformer heat conduction equation as a physical constraint into the model loss function, and to train the hot spot temperature prediction model using a deep reinforcement learning algorithm. The prediction output module is used to receive real-time alternating load data of the transformer, input it into the trained prediction model, and output the short-term predicted trajectory of the hot spot temperature. The aging assessment module is used to calculate the thermal aging coefficient based on the hot spot temperature prediction trajectory and generate graded early warning signals according to preset thresholds.

10. The dynamic prediction system for transformer hotspot temperature under alternating load based on deep reinforcement learning according to claim 9, characterized in that: The CFD simulation unit is used to establish a three-dimensional geometric model of the transformer, set boundary conditions for alternating loads at different frequencies, and simulate and generate oil flow field and temperature field data. The data acquisition unit is used to collect historical operating data and real-time alternating load data of the transformer through sensors; The data preprocessing unit is used to perform wavelet transform denoising, feature extraction, normalization, and outlier removal on simulation data and measured data; The model design module includes a TCN building unit and an attention mechanism building unit; The TCN building block is used to build a multi-level TCN network with causal dilated convolution, and to set residual connections and dilation factors. The attention mechanism building unit is used to design the weight calculation module and generate attention weight matrices for load features at different time steps; The model training module includes a physical constraint embedding unit and a training optimization unit; The physical constraint embedding unit is used to determine the heat conduction equation based on the transformer thermal capacity-thermal resistance model and construct a joint loss function that includes data fitting terms and physical residual terms. The training and optimization unit is used to train the model using the deep reinforcement learning backpropagation algorithm combined with the Adam optimizer, thereby updating and optimizing the network weights. The prediction output module adopts a sequence-to-sequence encoding-decoding structure. The encoding end extracts the temporal features of the real-time load data, and the decoding end outputs the predicted hotspot temperatures at different time steps in the future, forming a continuous temperature prediction trajectory.

Images