Main transformer load prediction method based on multi-source fusion data

Abstract

The invention discloses a main transformer load prediction method based on multi-source fusion data, and relates to the technical field of power equipment monitoring, and the method comprises the following steps: collecting various data, and carrying out the data preprocessing; performing stationarity test and pure randomness test on the original time sequence data, judging a current sequence change type, and performing corresponding feature extraction; establishing an autoregression model, a LightGBM model and a TiDE model, and fusing the three models by using a linear regression weighting method to obtain a main transformer load prediction model; establishing an operation risk assessment model, and performing risk assessment and early warning in combination with a load prediction result of the main transformer load prediction model; a linear regression weighting method is utilized to fuse the autoregression model, the LightGBM model and the TiDE model, so that the model prediction precision is improved, the heavy overload probability of the main transformer is effectively reduced, and the service life of the main transformer is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring and analysis, and in particular to a main transformer load prediction method based on multi-source fusion data. Background Art

[0002] According to the statistics of historical failure causes of main transformers, in addition to the aging of the main transformer, the main causes of failure are heavy load and overload of the main transformer, and the long-term overload operation of the equipment, which leads to failure or even equipment burnout. The existing main transformer load prediction methods mostly improve the accuracy of the prediction by considering the influencing factors from multiple angles, but due to the insufficient accuracy of the prediction model, large errors will be generated.

[0003] For example, Chinese patent CN114692937A, published on July 1, 2022, discloses a main transformer load prediction method based on the prediction of main transformer load influencing factors. The method is: construct a main transformer load influencing factor evaluation database, establish a main transformer load influencing factor evaluation model that considers typical time, power grid topology, power outage plan, main transformer parameters, meteorological information, main transformer overload, main transformer heavy load, main transformer load, and season, calculate the main transformer load data information and its model correlation coefficient, obtain the influencing factor data set according to the strong correlation evaluation factor, and classify the data set to obtain the predicted value of the influencing factor evaluation factor, and use the combined prediction algorithm of DTW and LSTM to perform fish and data training. The prediction model adopted by the present invention is not accurate enough. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the existing main transformer load prediction accuracy is low. A main transformer load prediction method based on multi-source fusion data is proposed, which uses the linear regression weighted method to fuse the autoregressive model, LightGBM model and TiDE model to improve the prediction accuracy.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a main transformer load prediction method based on multi-source fusion data, comprising the following steps: S1: Collect various data and perform data preprocessing; S2: Perform stationarity test and pure randomness test on the original time series data, determine the type of current sequence change, and perform corresponding feature extraction; S3: Establish autoregressive model, LightGBM model and TiDE model, and use linear regression weighting method to integrate the three models to obtain the main transformer load prediction model; S4: Establish an operation risk assessment model and conduct risk assessment and early warning based on the load forecast results of the main transformer load forecast model.

[0006] The main transformer load forecasting method based on multi-source fusion data, first of all, based on the multi-source measurement data of the power grid system, the measurement data is cleaned. On this basis, the time series decomposition method is introduced to analyze the load change characteristics of the main transformer and realize the load series stability analysis. Secondly, the time series feature engineering is constructed according to the time series prediction modeling method. Then, the main transformer load forecasting model is established based on the autoregressive model, lightweight gradient boosting machine and time series fully connected encoder neural network. In addition, the load-related influencing factors of the main transformer operation risk are studied, and the main transformer operation risk assessment model is established in combination with the load forecasting results to realize the quantification and early warning of the operation risk assessment, so that relevant personnel can deal with the risk situation in time.

[0007] Preferably, the autoregressive model is predicted by corresponding timestamp data and corresponding prediction variables. There are three parameters in the model. The optimal p and q parameters are determined by model fitting BIC, and a stable ARIMA calculation process is obtained after d times of difference. The evaluation standard of the model is that the lower the BIC value, the better.

[0008] Preferably, the LightGBM model uses a decision tree as a base learner to achieve main transformer load forecasting, uses the Optuna algorithm package for parameter optimization, and uses multi-fold cross validation for prediction. LightGBM is a gradient boosting framework that uses a decision tree as a base learner, which has the advantages of faster training speed, lower memory usage, support for parallel computing and GPU training, etc.

[0009] Preferably, the TiDE model obtains a residual block, stacks multiple residual blocks, substitutes static covariates and dynamic covariate parameters at the same time, and realizes time series prediction through encoding and decoding operations. The encoder and decoder of the model are both composed of multiple stacked residual blocks, and realize time series prediction through encoding and decoding operations.

[0010] Preferably, the data preprocessing in step S1 includes data outlier detection, and the specific contents are as follows: a robust Z-score algorithm is introduced to perform outlier detection to obtain the distance between the current data point and other remaining data points; the median and the absolute deviation of the median are used to correct the mean and standard deviation parameters. The traditional Z-score outlier detection algorithm is sensitive to extreme values, because the data mean itself is sensitive to extreme values. If there are abnormal extreme values, the mean value will deviate from the mean of the data distribution itself. Therefore, an outlier detection method based on a robust Z-score is introduced.

[0011] Preferably, in step S1, the time series data is used to predict the possible value of the current data value by using the past or future change trend, and the current missing value is filled and interpolated. The change trend of time series data will have a certain "inertia", and the past change trend will affect the future. Similarly, the size of future change data will also have a certain impact on the current data to a certain extent.

[0012] Preferably, the stationarity test in step S2 includes the following: performing time series decomposition by an additive model; and testing whether the current sequence is a stationary sequence by an ADF method. A time series can usually be composed of three parts: long-term trend fluctuations, seasonal changes, and random fluctuations. The decomposition model is divided into an additive model and a multiplicative model. Considering the improvement of the interpretability of time series decomposition, the additive model is selected as the time series decomposition model.

[0013] Preferably, the feature extraction in step S2 includes time type feature extraction and time lag feature extraction, and the extraction method is as follows: extract timestamp data, perform sin and cos calculations on hour and minute data as new feature inputs; extract whether the current date is a working day, introduce the Chinese calendar algorithm package to identify whether the current date is a statutory holiday; based on the measurement time of the time series data, shift the feature column downward by the corresponding time step to obtain the lag feature of the corresponding order. Since the timestamp type data of the measurement data cannot be directly input into the machine learning prediction model, the timestamp data is re-featured.

[0014] Preferably, the feature extraction in step S2 also includes operation risk feature extraction and sliding window feature extraction, and the operation risk features include the load rate of the main transformer, the equipment value and other load influencing factors; various statistical values ​​of the window are constructed by sliding the time window, and the strong correlation between the time series context data is extracted to construct the sliding window feature. The load rate of the main transformer refers to the ratio of the output power of the main transformer to the rated capacity of the main transformer. The equipment value feature is an inherent attribute of the transformer, including factors such as load level and equipment status. Other load influencing factors of the main transformer include the historical defect status of the main transformer, the operating years of the main transformer, and the historical heavy overload influencing factors.

[0015] Preferably, after the operation risk assessment model is established in step S4, the root mean square error and the mean absolute error are introduced as error evaluation indicators to calculate the error between the predicted value and the actual value of each model. The error between the predicted value and the actual value is calculated by the root mean square error and the mean absolute error, so as to judge the quality of the model.

[0016] The substantial effect of the present invention is as follows: the present invention considers the multi-dimensional influencing factors of the main transformer load change trend, utilizes the linear regression weighted method to fuse the autoregressive model, the LightGBM model and the TiDE model, improves the model prediction accuracy, effectively reduces the probability of main transformer heavy overload, and increases the operating life of the main transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a main transformer load prediction flow chart of Example 1; Figure 2 This is a timing decomposition diagram of the main transformer load data of Example 1; Figure 3 The main transformer load data histogram of the first embodiment; Figure 4 This is the ARIMA prediction result diagram of Example 1; Figure 5 This is the LightGBM model prediction result diagram of Example 1; Figure 6 This is the prediction result diagram of the fusion model of Example 1 Figure 7 This is a diagram of the TIDE model prediction results of Example 1. DETAILED DESCRIPTION

[0018] The specific implementation of the present invention will be further described below through specific examples in conjunction with the accompanying drawings.

[0019] Embodiment 1: The main transformer load prediction method based on multi-source fusion data, such as Figure 1 As shown in the figure, firstly, based on the multi-source measurement data of the power grid system, the measurement data is cleaned. On this basis, the time series decomposition method is introduced to analyze the load change characteristics of the main transformer and realize the load series stability analysis. Secondly, the time series feature engineering is constructed according to the time series prediction modeling method. Then, the main transformer load prediction model is established based on the autoregressive model (ARIMA), Light Gradient Boosting Machine (LightGBM) and Time-series Dense Encoder (TiDE). In addition, the load-related influencing factors of the main transformer operation risk are studied, and the main transformer operation risk assessment model is established in combination with the load forecasting results to realize the quantification and early warning of the operation risk assessment, so that relevant personnel can deal with the risk situation in time.

[0020] In the process of collecting various electrical quantities, the power system measurement equipment may cause abnormal values ​​and missing values ​​in the original data due to collection errors, communication interruptions or abnormalities, and data writing abnormalities. In response to the above problems, the following algorithm is proposed to deal with them: For the test of data outliers, a robust Z score algorithm is introduced to detect outliers. The Z score can be calculated to obtain the distance between the current data point and other remaining data points. The Z score is usually calculated as follows: In the formula, is the mean of vector x, and s is its standard deviation. However, the traditional Z-score outlier detection algorithm is sensitive to extreme values ​​because the data mean itself is sensitive to extreme values. If there are abnormal extreme values, the mean value will deviate from the mean of the data distribution itself. Therefore, a robust Z-score-based outlier detection method is introduced. Its calculation formula is: MAD=median(|X-median|) In the formula, X i is the i-th data of data vector X, and median represents the median. By using the median and the absolute deviation of the median to correct the mean and standard deviation parameters, the tolerance for abnormal extreme values ​​is enhanced, and the comprehensiveness and accuracy of outlier detection are improved.

[0021] For time series data, a data interpolation method based on the forward and backward weighted moving average method is proposed. Generally speaking, the change trend of time series data will have a certain "inertia". The past change trend will affect the future. Similarly, the size of future change data will also have a certain impact on the current data to a certain extent. Therefore, the past or future change trend can be used to predict the possible value of the current data value, thereby realizing the interpolation of the current missing value. The specific calculation formula is shown in Formula 4.

[0022] Where M tw is the weighted moving average of the N periods before and after, which is used as the interpolation value, w i For the corresponding y t The weight reflects its importance in the weighted average.

[0023] like Figure 2 As shown, this embodiment selects the load data of a main transformer in the jurisdiction from 04 / 23 to 06 / 02 as the verification data set to carry out data exploration and feature engineering construction. Before this process, data preprocessing has been completed.

[0024] For the original time series data, the first step of processing is to test the stationarity and pure randomness. According to the test results, the current sequence change type can be determined, and then the corresponding feature engineering can be extracted. If the mean of the time series does not change systematically, the variance does not change systematically, and the periodic changes are strictly eliminated, it can be called stable. Stationary time series are easier to predict with algorithm models, so for time series prediction, the stationarity of the series must be determined first. If it is not stable, it needs to be transformed into a stationary time series. Specifically, it can be analyzed and judged by two methods: time series decomposition and stationarity hypothesis testing.

[0025] Time series can usually be composed of three parts: long-term trend fluctuations, seasonal changes, and random fluctuations. Long-term trend fluctuations refer to a tendency or state in which a phenomenon continues to develop and change over a long period of time. Seasonal changes are regular changes in the level of development of a phenomenon due to changes in seasons. Random fluctuations are the combined effects of other factors in addition to long-term trend fluctuations and seasonal changes.

[0026] Decomposition models are divided into additive models and multiplicative models. The additive model means that the decomposition components are independent of each other and have the same dimension. The multiplicative model means that the output part and the trend term have the same dimension, the seasonal term is a proportional number, and the irregular change term is an independent random variable sequence that obeys the normal distribution. Considering the improvement of the interpretability of time series decomposition, the additive model is selected as the time series decomposition model.

[0027] The mainstream method for testing the stationary hypothesis is the unit root test, which tests whether there is a unit root in the sequence. If it exists, it is defined as a non-stationary sequence, and if it does not exist, it is a stationary sequence. Among them, the ADF test is one of the most commonly used unit root test methods.

[0028] The assumptions are as follows: Null hypothesis: H0: ρ = 1 (there is a unit root, the time series is non-stationary); Alternative hypothesis: H1: ρ < 1 (the time series is stationary). If the null hypothesis cannot be rejected (such as p> 0.05), the series is non-stationary, otherwise it is stationary. After testing, the p value is approximately equal to 0, which means that the load data is stable during this period of time.

[0029] like Figure 3 As shown in the figure, for the main transformer load time series data, a histogram can be used to reflect the corresponding load distribution. It can be seen from the figure that the main transformer active value mainly presents a "double peak" characteristic distribution, and most active values ​​are mainly distributed in the range of 10-25, with a certain left skew. Considering the load changes on weekends and weekdays, the loads of the two holidays are visualized by histograms. By distinguishing between holidays and weekdays, the distribution of load data changes greatly. For weekdays, the load data as a whole presents a right-skewed distribution with a larger mean, and for weekends, the load data presents a left-skewed distribution with a smaller mean. Therefore, when performing the next step of feature engineering, distinguishing date types can be used as one of the features to improve prediction accuracy.

[0030] The feature extraction of data mainly includes time type feature extraction, time lag feature extraction, sliding window feature extraction and operation risk feature extraction. The operation risk feature extraction is divided into main transformer load rate, equipment value, other load factors, etc.

[0031] Since the timestamp type data of the measurement data cannot be directly input into the machine learning prediction model, the timestamp data is re-featured. First, the year, month, day, hour, and minute data of the timestamp are extracted, and the sin and cos calculations are performed again on the hour and minute data as new feature inputs. Time series are periodic, and the introduction of trigonometric functions for calculation can well extract this feature, but a single sin or cos function can only reflect semi-periodicity, so the full periodicity is reflected by constructing sin and cos functions as two columns of features. In addition, the load changes of the power system are sensitive to information such as holidays, so it is also extracted whether the current date is a working day, and the Chinese calendar algorithm package is introduced to identify whether the current date is a statutory holiday.

[0032] An important characteristic of time series is that there is a strong correlation between the current or future predicted values ​​and the past historical values. For load forecasting, the load value at the current moment is similar to the load value at the previous moment. Considering the load change characteristics of the power system, the load will not change drastically at adjacent moments, but the overall trend on multiple time sections will show certain obvious changes. Therefore, the time series data measurement time can be used as a benchmark, and the feature column can be shifted downward by the corresponding time step to obtain the lag feature of the corresponding order for subsequent model prediction. The construction of the lag feature is shown in the following table (taking the third-order lag as an example). Time Measurement value Y First-order lag Second-order lag Third-order lag t1 Y1 t2 Y2 Y1 t3 Y3 Y2 Y2 t4 Y4 Y3 Y3 Y3

[0033] The lag correlation shows an overall decreasing trend with the increase of the lag period. Considering the size of the input feature dimension, the lag of the data at the same time point one day ago is selected as one of the input features for load forecasting.

[0034] The sliding window feature is similar to the lag feature. Unlike the lag feature that shifts the historical data downward, this method constructs various statistical values ​​of the window (such as mean, median, maximum, etc.) through a sliding time window, thereby extracting strong correlations between time series context data.

[0035] Based on the future load results calculated by the main transformer load prediction model, the future trend of the main transformer load rate can be calculated. The load rate of the main transformer reflects the health of the main transformer's operating status. If the load rate is too high, the load loss of the transformer will increase sharply, causing the transformer temperature to rise, which will greatly harm the operation and life of the transformer. Introducing the load rate indicator and carrying out the next step of modeling can improve the accuracy of the main transformer operation risk assessment. The load rate of the main transformer refers to the ratio of the main transformer output power to the rated capacity of the main transformer. The calculation formula is: load rate = actual active output / rated capacity of the main transformer.

[0036] The equipment value characteristics are the inherent properties of the transformer, including factors such as load level and equipment status. The higher the value of the main transformer equipment, the greater the impact of its abnormality or overload problems on the system. When estimating the risk of the main transformer, the value attributes of the equipment should be considered.

[0037] Other load influencing factors of the main transformer include the historical defect conditions of the main transformer, the operating years of the main transformer, the historical heavy overload influence factors, etc.

[0038] The full name of the ARIMA model is the autoregressive moving average model, which is the most common model used for time series prediction in statistical models. This model is relatively simple and only requires the endogenous variables of the time series without the help of other exogenous variables. Only the corresponding timestamp data and the corresponding predicted variable values ​​are required for prediction, which has a better effect on relatively stable time series. Usually the ARIMA (p, d, q) model has 3 parameters, indicating that it is a stable ARIMA process after d times of difference. The parameters of p and q can be determined based on the autocorrelation coefficient and partial autocorrelation coefficient, but the manual determination method has a certain degree of subjectivity, so the model fitting BIC is used to determine the optimal p and q.

[0039] BIC=kIn(n)-2In(L) Where k is the number of model parameters, n is the number of samples, and L is the likelihood function. Model evaluation criteria: The lower the BIC value, the better. The larger p and q are, the more parameters there are and the larger k is. So the smaller k is, the smaller p and q are, the better the model is. The larger L is, the smaller the BIC value is, so the larger the likelihood function is, the better.

[0040] The LightGBM model is one of the most commonly used and efficient machine learning models in the field of data mining, especially for data modeling of tabular data. It has been widely used in power system load forecasting, power system stability determination, fault prediction, reliability prediction and other problems. LightGBM is a gradient boosting framework that uses decision trees as base learners. Compared with the XGBoost model, it has the advantages of faster training speed, lower memory usage, support for parallel computing and GPU training. After completing the feature engineering feature extraction, the model is used to realize the main transformer load forecast. For the selection of various hyperparameters of the model, the Optuna algorithm package is used for parameter optimization, and multi-fold cross validation is used for prediction.

[0041] The TiDE model is the latest linear time series prediction model proposed by Google. This model upgrades the traditional MLP network and once again defeats Transformer (one of the basic models of chatGPT). In solving complex time series problems, there is a problem of being unable to adapt to the nonlinearity between time series and other time series variables. Based on these problems, the TiDE model upgrades the multi-layer perceptron time series prediction model and has achieved the best prediction results on multiple public data sets.

[0042] The TiDE model not only utilizes past sequence values ​​(backtracking operation), but also utilizes some covariate information, such as static covariates of time series and dynamic covariates that are known at any time. Among them, static covariates refer to variables that do not change over time, and dynamic covariates refer to variables that change over time but are still known in the future, such as date parameters. The most basic module in the model, the "residual block", draws on the ResNet architecture. It is essentially an MLP with Dropout, ReLU nonlinearity, residual connection, and LayerNorm. There is a global residual connection on the far left, and only one linear layer is used. The encoder and decoder of the model are both made up of multiple stacked residual blocks, and time series prediction is achieved through encoding and decoding operations.

[0043] In the field of machine learning prediction problems, the prediction results of a single model do not have better robustness. Because the final prediction result is determined by only a single model, the accuracy of the final prediction result will be "largely volatile", so multiple models are introduced for prediction in the above text, but how to better integrate and fuse the final prediction results is a new problem. Here, the linear regression weighted method is used to perform weighted fusion on the prediction results of multiple models. For the linear regression machine learning model, its principle can be expressed as: In the formula, ω1, ω2 and ω3 represent the weights of the predicted values ​​of the three models respectively. The optimal weight is determined by minimizing the square error between the true value and the predicted value, and the fusion of the three models is achieved to improve the accuracy and robustness of the prediction results.

[0044] Therefore, the prediction results of the three prediction models mentioned above are input into the linear regression learner for training, and the final prediction result weights of each model can be obtained.

[0045] Combined with the power system risk analysis theory, a main transformer risk assessment model is established. Based on this model, risk points are predicted in advance to ensure timely elimination of main transformer faults. A risk assessment model based on the main transformer load condition is constructed. The specific calculation method is as follows: R=S×H×D H=π i f D=I×L I1=10P m Where R is the load risk value of the main transformer; S is the load factor, which is the integral of the load rate α(t) on the specified dimension time section; H is the load-related factor, which is other risk factors that cause the main transformer operation. f The product of; D is the consequence severity value; I is the importance of the main variable; L is the degree of loss; is the weight of the equipment importance factor, i=1, 2, 3 represent the equipment value, load level and equipment status respectively; P m is the equipment cost; L represents the degree of loss; Represents the weight of loss degree, j=1, 2, 3 represent equipment cost, personal safety and power safety respectively.

[0046] like Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown in the figure, according to the feature engineering and modeling methods proposed above, the validation data set is calculated and trained accordingly, and the load data from June to February is selected as the test set to verify the model. The performance of the prediction indicators of each model is shown in the following table. Metric ARIMA LGB TiDE Fusion Model MAE 2.16 1.870 1.580 0.736 RMSE 2.53 2.23 1.906 1.001

[0047] The above comparative analysis shows that among the single model performances, the TiDE model performs best and achieves the lowest values ​​in both evaluation indicators. After adopting the fusion model algorithm, both indicators achieve the best results. Therefore, the introduction of the multi-model linear fusion concept can improve the prediction robustness of the model and further improve the accuracy of the main transformer load prediction.

[0048] Considering the main transformer as a power station, the load level is 1000, the equipment status is 800, it was put into operation in May 2009, and the rated capacity is 50000kVA. According to the calculation results of the load forecasting model, the load rate change trend of the main transformer in the future specified time can be calculated. The calculation shows that the load rate of the main transformer has been at a relatively low level. Due to the long operation time, the aging factor is 1.4; due to the low historical load level of the equipment, the historical heavy overload factor is also small, and the historical heavy overload factor is 1.1, which is a small value; the voltage level is 110kV, and the consequence severity is 1.2.

[0049] According to the risk classification in the above table, the comprehensive risk value of the transformer is calculated to be 29.54, which is less than the low-medium risk threshold of 50, so the equipment is judged to be low risk. Taking into account the risk characteristics of the equipment, it has a long service life, but its load changes should be paid attention to during operation, as the risk may increase rapidly due to the increase in load.

[0050] This patent considers the multi-dimensional influencing factors of the main transformer load change trend, proposes a main transformer load prediction model, realizes accurate prediction of the main transformer load, and provides auxiliary decision-making support for various departments. According to the main transformer load prediction results, combined with the main transformer operation risk influencing factors, a main transformer operation risk assessment model is established to achieve early perception of the main transformer operation risk and help relevant personnel take timely treatment measures. It can effectively save manpower costs such as inspections of substation departments, reduce the probability of heavy overload of main transformers, and improve the operating life of main transformers and the reliability of power grid operation, providing a digital application foundation for smoothly meeting the peak summer and helping to ensure power supply.

[0051] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. The main transformer load forecasting method based on multi-source fusion data is characterized by: The following steps are involved: S1: Collect various data and perform data preprocessing; S2: Perform stationarity test and pure randomness test on the original time series data, determine the type of current sequence change, and perform corresponding feature extraction; S3: Establish autoregressive model, LightGBM model and TiDE model, and use linear regression weighting method to integrate the three models to obtain the main transformer load prediction model; S4: Establish an operation risk assessment model and conduct risk assessment and early warning based on the load forecast results of the main transformer load forecast model.

2. The main transformer load forecasting method based on multi-source fusion data according to claim 1 is characterized in that: The autoregressive model is predicted by the corresponding timestamp data and the corresponding prediction variables. There are three parameters in the model. The optimal p and q parameters are determined by model fitting BIC. After d times of difference, a stable ARIMA calculation process is obtained.

3. The main transformer load forecasting method based on multi-source fusion data according to claim 2 is characterized in that: The LightGBM model uses a decision tree as a base learner to achieve main transformer load forecasting, uses the Optuna algorithm package for parameter optimization, and uses multi-fold cross validation for prediction.

4. The main transformer load forecasting method based on multi-source fusion data according to claim 3 is characterized in that: The TiDE model obtains a residual block, stacks multiple residual blocks, substitutes static covariate and dynamic covariate parameters at the same time, and realizes time series prediction through encoding and decoding operations.

5. The main transformer load forecasting method based on multi-source fusion data according to claim 1 is characterized in that: The data preprocessing in step S1 includes data outlier detection, and the specific contents are as follows: a robust Z-score algorithm is introduced to perform outlier detection to obtain the distance between the current data point and other remaining data points; the median and the absolute deviation of the median are used to correct the mean and standard deviation parameters.

6. The main transformer load forecasting method based on multi-source fusion data according to claim 5 is characterized in that: In step S1, the possible value of the current data value is predicted by using the past or future change trend of the time series data, and the current missing value is filled and interpolated.

7. The main transformer load forecasting method based on multi-source fusion data according to claim 1 is characterized in that: The stationarity test in step S2 includes the following contents: performing time series decomposition through an additive model; and testing whether the current sequence is a stationary sequence through the ADF method.

8. The main transformer load forecasting method based on multi-source fusion data according to claim 1 is characterized in that: The feature extraction in step S2 includes time type feature extraction and time lag feature extraction, and the extraction method is as follows: extract timestamp data, perform sin and cos calculations on hour and minute data as new feature inputs; extract whether the current date is a working day, and introduce the Chinese calendar algorithm package to identify whether the current date is a statutory holiday; based on the measurement time of the time series data, shift the feature column downward by the corresponding time step to obtain the lag feature of the corresponding order.

9. The main transformer load forecasting method based on multi-source fusion data according to claim 8 is characterized in that: The feature extraction in step S2 also includes operation risk feature extraction and sliding window feature extraction. The operation risk features include the main transformer load rate, equipment value and other load influencing factors. Various statistical values ​​of the window are constructed by sliding the time window, and the strong correlation between the time series context data is extracted to construct the sliding window feature.

10. The main transformer load forecasting method based on multi-source fusion data according to claim 1, 2, 3, 4 or 5, characterized in that: After the operation risk assessment model is established in step S4, the root mean square error and mean absolute error are introduced as error evaluation indicators to calculate the errors between the predicted values ​​and the actual values ​​of each model.

Images