A transformer oil temperature prediction method and system based on an LSTM neural network

Abstract

The application discloses a transformer oil temperature prediction method and system based on an LSTM neural network, and comprises the following steps: obtaining historical data of transformer ambient meteorology, operating reactive power and active power and transformer oil temperature, and obtaining a historical data set through preprocessing; constructing time sequence characteristics, polynomial characteristics and statistical characteristics according to the obtained historical data set, and obtaining a transformer oil temperature prediction sample set; dividing the sample set into a training set and a test set, then performing feature division to improve the information capturing capability of the model on the prediction period; and finally constructing an LSTM time sequence prediction model according to the feature division and the training set. The application further captures the time sequence law of the transformer oil temperature change, improves the accuracy of the oil temperature prediction, avoids transformer damage caused by overheating, and improves the equipment reliability.

Description

Technical Field

[0001] This invention relates to transformer oil temperature prediction technology, specifically to a prediction method and system based on adaptive bus load conditions. Background Technology

[0002] Power transformers are key equipment in power grids, and transformer oil temperature prediction is one of the key tasks in power system monitoring and maintenance. Transformers are an important component of the power system, and their normal operation and reliability are closely related to oil temperature. The load capacity and insulation aging degree of power transformers are both affected by the transformer's thermal characteristics. Accurate transformer oil temperature prediction is of great significance for inspecting the operating status of power transformers, judging the degree of insulation aging, and predicting faults.

[0003] Existing methods for predicting the oil temperature of power transformers include numerical calculation, thermal circuit modeling, and traditional machine learning algorithms. Numerical calculation yields relatively accurate results but is computationally intensive and time-consuming, making it unsuitable for online monitoring of power transformer oil temperature. Thermal circuit modeling often uses empirical values ​​for relevant heat transfer parameters (which frequently change dynamically with temperature), resulting in low prediction accuracy. Traditional machine learning algorithms struggle to fit the intrinsic relationships within the data, and the selection of input features is time-consuming and labor-intensive. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a transformer oil temperature prediction method and system based on LSTM neural network, which improves the accuracy of transformer oil temperature prediction.

[0005] Technical Solution: This invention provides a transformer oil temperature prediction method based on LSTM neural networks. Based on historical and predicted data of transformer environmental meteorology, operating reactive power, operating active power, and transformer oil temperature, the method uses an LSTM oil temperature prediction model to predict the transformer oil temperature. The construction process of the LSTM oil temperature prediction model includes:

[0006] Historical data on transformer environmental weather, operating reactive power, operating active power, and transformer oil temperature are acquired and preprocessed to obtain a historical dataset of the transformer.

[0007] Based on the obtained historical data set of transformers, time-series features, multinomial features, and statistical features are analyzed. The time-series features are characteristic factors obtained by analyzing the periodic variation patterns of historical transformer oil temperature data; the multinomial features are characteristic factors obtained by analyzing the nonlinear variation patterns of historical transformer oil temperature data; and the statistical features are characteristic factors obtained by analyzing the variation trends of historical transformer oil temperature data. Based on the obtained set of characteristic factors, the correlation and importance of each characteristic factor with the oil temperature data are calculated. Feature selection is then performed based on the correlation and importance, and the historical data and predicted data corresponding to the selected characteristic factors constitute a predicted sample set for transformer oil temperature.

[0008] The transformer oil temperature prediction sample set is divided into a test set and a training set. The LSTM oil temperature prediction model is trained using the training set. During model training, the features in the training set are divided into a historical known feature factor dataset and a future known feature factor dataset. The historical known feature factor data refers to the feature factor data before the prediction time, and the future known feature factor data refers to the feature factor data after the prediction time. Dependency information is learned for the historical known feature factor data and the future known feature factor data respectively through an LSTM neural network. The outputs of the two LSTM neural networks are merged and then output as oil temperature data through a fully connected layer. The LSTM oil temperature prediction model is tested using the test set.

[0009] The selected characteristic factors include: the Fourier transform coefficient of the transformer oil temperature in the previous week.

[0010] The preprocessing includes: detecting out-of-limit values ​​in the data; detecting abrupt outliers in the data; detecting continuous constant values ​​in the data; deleting detected outliers; repairing missing values ​​using interpolation; and smoothing using a moving average. Continuous constant values ​​in the data include: calculating the standard deviation for each category of data; discarding data of the current category if its standard deviation is close to zero; setting a threshold for change in all categories of data based on historical data analysis; marking the current point as a constant value outlier if the change is less than the threshold; setting a threshold for the number of continuous constant values ​​in all categories of data based on historical data analysis; counting the number of consecutive constant value data points; marking data points with more than the set number of consecutive constant values ​​as continuous constant value outliers, and removing the outlier marking for other points.

[0011] Feature selection based on correlation and feature importance includes the following steps: calculating the Person correlation coefficient between each feature factor data and the oil temperature data; constructing a decision tree model with the feature factor data as input and the oil temperature data as output, and calculating the importance of each feature factor based on the Feature Importance in the decision tree model; ranking the feature factors according to the Person correlation coefficient and feature importance, and performing feature selection based on the ranking.

[0012] The LSTM oil temperature prediction model is trained using the training set, including: randomly sampling the training set proportionally, constructing a validation set using the sampled samples, and removing it from the training set; training the oil temperature prediction model, evaluating it on the validation set using the EarlyStop method, and saving the model that performs best on the validation set.

[0013] The LSTM oil temperature prediction model is tested using a test set, including the following steps: inputting the corresponding data of the test set into the trained LSTM oil temperature prediction model and aligning them according to time sequence; averaging the prediction points with repeated time sequences in the prediction results, retaining only the mean of the predictions at the same time; and calculating the model prediction accuracy using the Mape metric.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned transformer oil temperature prediction method based on LSTM neural network.

[0015] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned transformer oil temperature prediction method based on an LSTM neural network.

[0016] This invention provides a time-series prediction system for transformer oil temperature, comprising:

[0017] The data acquisition module is used to store historical and predicted data on transformer environmental weather, operating reactive power, operating active power, and transformer oil temperature.

[0018] The data preprocessing module is used to preprocess the data;

[0019] The time-series feature construction module is used to analyze time-series features, multinomial features, and statistical features based on historical datasets. The time-series features are feature factors obtained by analyzing the periodic variation patterns of historical transformer oil temperature data; the multinomial features are feature factors obtained by analyzing the nonlinear variation patterns of historical transformer oil temperature data; and the statistical features are feature factors obtained by analyzing the variation trends of historical transformer oil temperature data.

[0020] The feature filtering module is used to construct a set of feature factors based on the time-series feature building module, calculate the correlation and feature importance of each feature factor with oil temperature data, and perform feature filtering based on the correlation and feature importance.

[0021] The dataset partitioning module is used to divide the transformer oil temperature prediction sample set into a historical known feature factor dataset and a future known feature factor dataset. The historical known feature factor data is the feature factor data before the prediction time, and the future known feature factor data is the feature factor data after the prediction time. The transformer oil temperature prediction sample set includes the historical data and prediction data corresponding to the selected feature factors.

[0022] The oil temperature prediction module includes two LSTM neural networks, which learn dependency information from historical known feature factor data and future known feature factor data, respectively. The outputs of the two LSTM neural networks are merged and then output as oil temperature data through a fully connected layer.

[0023] Beneficial Effects: Compared with existing technologies, this invention has the following advantages: This invention introduces a temporal neural network from deep learning to predict transformer oil temperature. LSTM can effectively capture long-distance dependencies in time series data because it can directly establish relationships between various time steps without being limited by the sequence length. This allows the use of time series prediction methods to further capture the temporal patterns of transformer oil temperature changes, improving the accuracy of oil temperature prediction. By analyzing the temporal features, multinomial features, and statistical features of historical data, a wider range of feature factors can be obtained, improving the model learning effect. Through improvements to the model structure, two LSTM models are introduced to learn from historical and predicted data simultaneously. This means that oil temperature prediction no longer solely relies on the input real-time data, but rather obtains more accurate prediction results by combining historical and predicted values. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the transformer oil temperature prediction method based on LSTM neural network described in this invention.

[0025] Figure 2 This is a flowchart of the data cleaning process described in this invention;

[0026] Figure 3 This is a diagram showing the data cleaning results described in this invention;

[0027] Figure 4 This is a diagram of the model structure described in this invention;

[0028] Figure 5 This is a diagram showing the prediction results described in this invention;

[0029] Figure 6 This is a schematic diagram of the time-series prediction system for transformer oil temperature according to the present invention. Detailed Implementation

[0030] The technical solution of the present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0031] like Figure 1 As shown, the transformer oil temperature prediction method based on LSTM neural network of the present invention includes the following steps:

[0032] S1. Obtain historical data on transformer environmental weather, reactive power, active power, and transformer oil temperature, and align and splice them according to time sequence.

[0033] In this example, meteorological data, transformer power data, and transformer oil temperature data of a transformer were obtained from August 1, 2022 to June 15, 2023. The relevant weather data included temperature, wind speed, humidity, etc., with one sample point every 15 minutes.

[0034] S2. Based on the dataset obtained in step S1, preprocess the historical oil temperature data, historical meteorological data, historical reactive power data, and historical active power data of the transformer to obtain the historical dataset of the transformer. For example... Figure 2 As shown, the specific steps include:

[0035] S2.1 performs time-series completion on the spliced ​​data, filling in any missing data with null values.

[0036] S2.2 Based on the statistical characteristics of the data, the historical data of the transformer is limited by the highest and lowest values ​​in the historical data, and the out-of-limit values ​​in the historical data are detected;

[0037] S2.3 Based on the statistical characteristics of the data, the 3 Sigma detection method is used to detect abrupt outliers in the historical data of transformers, including the following steps:

[0038] (1) For each sample x∈X in the data, check whether μ-3σ≤x≤μ+3σ is satisfied, where μ is the sample mean, and its calculation formula is:

[0039]

[0040] Where n represents the number of samples; X i Let represent the value of the i-th sample; σ represents the sample standard deviation; the formula for calculating the sample standard deviation σ is:

[0041]

[0042] Where n represents the number of samples; Xi Represents the value of the i-th sample; μ represents the mean of the current data samples;

[0043] (2) The historical oil temperature data, historical meteorological data, historical reactive power data and historical active power data of the transformer are subjected to 3sigma detection. Data that does not meet the 3sigma condition is recorded as a sudden change anomaly.

[0044] S2.4 uses a logic-based detection method to detect continuous constant values ​​in historical data, including the following steps:

[0045] (1) For all categories of data, calculate the standard deviation of each category of data. If the standard deviation of the current category of data is close to zero, discard the current category of data.

[0046] (2) Based on historical data analysis, set a threshold for changes in all categories of data. If the magnitude of the change is less than the threshold, mark the current point as a constant value abnormal data point.

[0047] (3) Based on historical data analysis, set a threshold for the number of consecutive constant values ​​for all categories of data, count the number of consecutive constant value data points, mark the constant value data points that exceed the set number of values ​​as consecutive constant value abnormal data points, and cancel the abnormal marking for other points.

[0048] In this example, if the number of consecutive constant values ​​exceeds 8, it is considered an outlier; if it is 8 or less, it is considered normal data.

[0049] S2.5, based on historical data change patterns, uses methods such as linear interpolation, statistical interpolation, and bilinear interpolation to repair detected outliers and missing values. This includes the following steps:

[0050] (1) Set the outlier to null, and retain its corresponding time series and position in the dataset.

[0051] (2) For outliers with fewer than 4 consecutive outliers, use a linear interpolation algorithm to repair them.

[0052] (3) Outliers with more than 4 but less than 96 consecutive outliers are repaired using bilinear interpolation. The calculation formula is as follows:

[0053]

[0054]

[0055]

[0056] Where (x,y) is the oil temperature point to be interpolated, (x1,y1) is the previous known oil temperature point yesterday, (x1,y2) is the next known oil temperature point yesterday, (x1,y2) is the previous known oil temperature point tomorrow, (x2,y2) is the next known oil temperature point tomorrow, and f(x,y) is the oil temperature value at point (x,y).

[0057] (4) Outliers with more than 96 consecutive outliers are repaired using a statistical interpolation algorithm. The interpolation method is shown in the table below:

[0058] Table 1 Data Repair Methods

[0059] 0<n≤96 Bilinear interpolation 96<n≤192 Linear interpolation at the same time point n>192 Mean interpolation at the same time point

[0060] S2.5 uses a moving average method to smooth the historical data, obtaining the historical dataset of the transformer. The data cleaning results are shown in the figure below. Figure 3 As shown.

[0061] S3 constructs time-series features, multinomial features, and statistical features based on the obtained historical data of the transformer to obtain a predicted sample set for transformer oil temperature, including the following steps:

[0062] S3.1 Analyzes the periodicity of historical oil temperature data, constructs multi-dimensional time series features, fits the periodic variation law of transformer oil temperature, and obtains feature factors including: Sin_day, Sin_hour, FFT_coefficient, ...;

[0063] S3.2 analyzes the nonlinear changes in historical oil temperature data, constructs polynomial features, and expands the model's fitting ability in high-dimensional space. The polynomial features are generated by performing arithmetic operations on pairwise data features to obtain feature factors including: Tmp_plus_windspeed, Tmp_mutiply_windspeed, Tmp_subtract_windspeed, ...

[0064] S3.3 analyzes the changing trend of historical oil temperature data, constructs statistical features, fits the future trend of transformer oil temperature, and obtains feature factors including: Day_max_tmp, Day_min_tmp, Day_max_humid, ...

[0065] S4 calculates the correlation and importance of each feature with the oil temperature data, and performs feature filtering based on the correlation and importance, including the following steps:

[0066] S4.1 Calculate the Person correlation coefficient between each feature and the oil temperature data based on the generated features and historical oil temperature data;

[0067] S4.2 Based on the generated features and historical oil temperature data, a decision tree model is constructed with features as input and oil temperature as output. The importance of each feature is calculated based on the Feature Importance in the decision tree model.

[0068] S4.3 Based on the Person correlation coefficient and feature importance, sort the features and select those with high correlation and high importance.

[0069] The features obtained after feature filtering in this example are shown in the table below:

[0070] Table 2 Feature Filtering Results

[0071]

[0072]

[0073] S5 uses the transformer oil temperature prediction sample set to divide features according to different data properties, transforms the dataset according to the features, and segments it according to time series to obtain the training set and test set for transformer oil temperature prediction. This includes the following steps:

[0074] S5.1 calculates the maximum and minimum values ​​of the predicted sample set of transformer oil temperature, and normalizes the sample set using the maximum and minimum values;

[0075] S5.2 The predicted sample set of transformer oil temperature is classified according to whether it can be obtained at the time of prediction, into historical known variables, future known variables and prediction targets, and the dataset is transformed according to the classified variables and their corresponding features.

[0076] S5.3 The transformed dataset is split in chronological order to create a training set for training the oil temperature prediction model and a test set for evaluating the oil temperature prediction model.

[0077] In this example, the oil temperature dataset from August 1, 2022 to June 1, 2023 is divided into a training set and a test set from June 1, 2023 to June 14, 2023. In the test set, at each time point, the oil temperature 10 hours after the current time needs to be predicted, that is, at each time point, 40 points after the current time are predicted.

[0078] S6 constructs an LSTM time series prediction model based on the aforementioned feature segmentation and training set. The model structure diagram is shown below. Figure 4 As shown. Training is performed using the training set, including the following steps:

[0079] S6.1 Based on the feature division, two LSTM blocks are constructed to extract the long-term dependency information of historical and future known variables, and the input and output structure of the model is constructed according to the prediction target.

[0080] The predicted sample set of transformer oil temperature is divided into a historical known feature factor dataset and a future known feature factor dataset, based on whether the data at the prediction time is known. The historical known feature factor data refers to the feature factor data before the prediction time (known), such as Day-1, Day-2, Day-1_max_tmp, ...; the future known feature factor data refers to the feature factor data after the prediction time (unknown, belonging to the predicted data), such as Day_max_tmp, Day_min_tmp, Tmp, ... Dependency information is learned for both the historical and future known feature factor data using LSTM neural networks. The outputs of the two LSTM neural networks are then merged and output as oil temperature data through a fully connected layer.

[0081] S6.2 Based on the divided training set, randomly sample the training set according to the proportion, use the sampled samples to construct the validation set, and delete it from the training set;

[0082] S6.3 Based on the training set, validation set, and constructed model, train the oil temperature prediction model, evaluate it on the validation set using the Early Stop method, and save the model that performs best on the validation set.

[0083] S7 inputs the data corresponding to the test set into the LSTM oil temperature prediction model to obtain the prediction result of the transformer oil temperature. The prediction result is shown in the figure below. Figure 5 As shown, and to calculate the accuracy, the following steps are included:

[0084] S7.1 Input the data corresponding to the test set into the trained LSTM oil temperature prediction model and align them according to time sequence;

[0085] S7.2 Mean the prediction points that are repeated in time sequence in the prediction results, and only retain the mean of the prediction at the same time.

[0086] S7.3 uses the Mape metric to calculate the model's prediction accuracy.

[0087] Based on the same inventive concept, this invention also includes a time-series prediction system for transformer oil temperature, the system architecture of which is as follows: Figure 6 As shown, it includes:

[0088] The data acquisition module is used to store historical and predicted data on transformer environmental weather, operating reactive power, operating active power, and transformer oil temperature.

[0089] The data preprocessing module is used to preprocess the data;

[0090] The time-series feature construction module is used to analyze time-series features, multinomial features, and statistical features based on historical datasets. The time-series features are feature factors obtained by analyzing the periodic variation patterns of historical transformer oil temperature data; the multinomial features are feature factors obtained by analyzing the nonlinear variation patterns of historical transformer oil temperature data; and the statistical features are feature factors obtained by analyzing the variation trends of historical transformer oil temperature data.

[0091] The feature filtering module is used to construct a set of feature factors based on the time-series feature building module, calculate the correlation and feature importance of each feature factor with oil temperature data, and perform feature filtering based on the correlation and feature importance.

[0092] The dataset partitioning module is used to divide the transformer oil temperature prediction sample set into a historical known feature factor dataset and a future known feature factor dataset. The historical known feature factor data is the feature factor data before the prediction time, and the future known feature factor data is the feature factor data after the prediction time. The transformer oil temperature prediction sample set includes the historical data and prediction data corresponding to the selected feature factors.

[0093] The oil temperature prediction module includes two LSTM neural networks, which learn dependency information from historical known feature factor data and future known feature factor data, respectively. The outputs of the two LSTM neural networks are merged and then output as oil temperature data through a fully connected layer.

[0094] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described transformer oil temperature prediction method based on an LSTM neural network.

[0095] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described LSTM neural network-based transformer oil temperature prediction method.

[0096] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. A method for predicting transformer oil temperature based on LSTM neural network, characterized in that, Based on historical and predicted data of transformer environmental meteorology, operating reactive power, operating active power, and transformer oil temperature, an LSTM oil temperature prediction model is used to predict transformer oil temperature. The construction process of the LSTM oil temperature prediction model includes: Historical data on transformer environmental weather, operating reactive power, operating active power, and transformer oil temperature are acquired and preprocessed to obtain a historical dataset of the transformer. Based on the obtained historical data set of transformers, time-series features, multinomial features, and statistical features are analyzed. The time-series features are characteristic factors obtained by analyzing the periodic variation patterns of historical transformer oil temperature data; the multinomial features are characteristic factors obtained by analyzing the nonlinear variation patterns of historical transformer oil temperature data; and the statistical features are characteristic factors obtained by analyzing the variation trends of historical transformer oil temperature data. Based on the obtained set of characteristic factors, the correlation and importance of each characteristic factor with the oil temperature data are calculated. Feature selection is then performed based on the correlation and importance, and the historical data and predicted data corresponding to the selected characteristic factors constitute a predicted sample set for transformer oil temperature. The transformer oil temperature prediction sample set is divided into a test set and a training set. The LSTM oil temperature prediction model is trained using the training set. During model training, the features in the training set are divided into a historical known feature factor dataset and a future known feature factor dataset. The historical known feature factor data refers to the feature factor data before the prediction time, and the future known feature factor data refers to the feature factor data after the prediction time. Dependency information is learned for the historical known feature factor data and the future known feature factor data respectively through an LSTM neural network. The outputs of the two LSTM neural networks are merged and then output as oil temperature data through a fully connected layer. The LSTM oil temperature prediction model is tested using the test set.

2. The transformer oil temperature prediction method based on LSTM neural network according to claim 1, characterized in that, The selected characteristic factors include: the Fourier transform coefficient of the transformer oil temperature in the previous week.

3. The transformer oil temperature prediction method based on LSTM neural network according to claim 1, characterized in that, The preprocessing includes: Values ​​exceeding the limit in the detection data; Detect mutation outliers in the data; Continuous constant values ​​in the detection data; Detected outliers are deleted, missing values ​​are repaired using interpolation, and smoothing is performed using a moving average method.

4. The transformer oil temperature prediction method based on LSTM neural network according to claim 2, characterized in that, Continuous constant values ​​in the detection data include: For all categories of data, calculate the standard deviation for each category. If the standard deviation of the current category is close to zero, discard the data for that category. Based on historical data analysis, a threshold for change is set for all categories of data. If the change is less than the threshold, the current point is marked as a constant value outlier. Based on historical data analysis, a threshold is set for the number of consecutive constant values ​​in all categories of data. The number of consecutive constant value data points is counted. Constant value data points that exceed the set number are marked as abnormal data points, while other points are unmarked as abnormal.

5. The transformer oil temperature prediction method based on LSTM neural network according to claim 1, characterized in that, Feature selection based on relevance and feature importance includes the following steps: Calculate the Person correlation coefficient between each characteristic factor data and the oil temperature data; A decision tree model is constructed using feature factor data as input and oil temperature data as output. The importance of each feature factor is calculated based on the Feature Importance in the decision tree model. Based on the Person correlation coefficient and feature importance, the feature factors are ranked, and feature selection is performed based on the ranking.

6. The transformer oil temperature prediction method based on LSTM neural network according to claim 1, characterized in that, The LSTM oil temperature prediction model was trained using the training set, including: The training set is randomly sampled proportionally, and the sampled samples are used to construct a validation set, which is then removed from the training set. The oil temperature prediction model is trained and evaluated on the validation set using the Early Stop method. The model that performs best on the validation set is saved.

7. The transformer oil temperature prediction method based on LSTM neural network according to claim 1, characterized in that, The LSTM oil temperature prediction model was tested using a test set, including the following steps: The test set data is input into the trained LSTM oil temperature prediction model and aligned according to time sequence. For prediction points with temporal repetition in the prediction results, average the predictions and retain only the average value at the same time. The Mape metric is used to calculate the model's prediction accuracy.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the transformer oil temperature prediction method based on LSTM neural network as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the transformer oil temperature prediction method based on LSTM neural network as described in any one of claims 1 to 7.

10. A time-series prediction system for transformer oil temperature, characterized in that, include: The data acquisition module is used to store historical and predicted data on transformer environmental weather, operating reactive power, operating active power, and transformer oil temperature. The data preprocessing module is used to preprocess the data; The time-series feature construction module is used to analyze time-series features, multinomial features, and statistical features based on historical datasets. The time-series features are feature factors obtained by analyzing the periodic variation patterns of historical transformer oil temperature data; the multinomial features are feature factors obtained by analyzing the nonlinear variation patterns of historical transformer oil temperature data; and the statistical features are feature factors obtained by analyzing the variation trends of historical transformer oil temperature data. The feature filtering module is used to construct a set of feature factors based on the time-series feature building module, calculate the correlation and feature importance of each feature factor with oil temperature data, and perform feature filtering based on the correlation and feature importance. The dataset partitioning module is used to divide the transformer oil temperature prediction sample set into a historical known feature factor dataset and a future known feature factor dataset. The historical known feature factor data is the feature factor data before the prediction time, and the future known feature factor data is the feature factor data after the prediction time. The transformer oil temperature prediction sample set includes the historical data and prediction data corresponding to the selected feature factors. The oil temperature prediction module includes two LSTM neural networks, which learn dependency information from historical known feature factor data and future known feature factor data, respectively. The outputs of the two LSTM neural networks are merged and then output as oil temperature data through a fully connected layer.

Images