Statistical method for predicting overload probability and correction of similar day load rate

Abstract

The present application belongs to the technical field of load prediction, and particularly relates to a distribution transformer load rate correction prediction method based on heavy overload probability statistics and similar days. The present application obtains time points with heavy overload possibility by statistically calculating the probability of heavy overload of load rate at each time point in the recent period, focuses on correcting the time points with heavy overload possibility, and improves the correction efficiency. Considering the influence of short-term external factors on the load rate peak value, the load rate similar day of the prediction day is found with the aid of the load rate curve of the day before the prediction day, and then the initial predicted load rate curve is corrected at the time points with heavy overload probability according to the load rate peak value height of the day before the prediction day and the similar day, thereby improving the prediction accuracy of the heavy overload part and solving the problem of heavy overload misjudgment.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, specifically involving a method for correcting and forecasting the load rate of distribution transformers based on heavy overload probability statistics and similar days. Background Technology

[0002] In recent years, with rapid socio-economic development, people's electricity consumption for production and daily life has increased significantly, especially during the peak summer air conditioning season. In densely populated areas and large industrial parks, daily peak electricity consumption is relatively concentrated, easily leading to heavy or even overloaded distribution transformers. Distribution transformers operating under heavy overload conditions can cause transformer failures, causing great inconvenience to residents, affecting production efficiency, and resulting in huge economic losses. To avoid heavy overload of distribution transformers and ensure the stable operation of distribution lines, risk warnings for potentially overloaded distribution transformers are particularly important.

[0003] Currently, heavy overload early warning work is mainly based on predicting the distribution transformer load rate. This involves analyzing historical load rate fluctuation patterns, incorporating factors such as temperature that may directly or indirectly affect the load rate, and using various machine learning methods to construct a load rate prediction model. Finally, the heavy overload situation is assessed based on the load rate prediction results. In the field of machine learning, time series methods are widely used in load rate prediction. The Autoregressive Integrated Moving Average (ARIMA) model is a relatively mature algorithm. ARIMA establishes a time series model of historical load rate fluctuations over time. By learning from historical data, the model predicts future load rates. Another commonly used classic prediction method is regression analysis, which learns the relationship between independent and dependent variables and establishes a regression equation based on historical data. Time series methods and regression analysis are the mainstream techniques for load rate prediction. However, due to some short-term external factors, traditional machine learning models have difficulty accurately predicting load rate peaks and providing early warnings of heavy overload, resulting in many potential heavy overload risks being missed. Furthermore, the load rate prediction results are corrected using fixed coefficients, which lacks specificity for heavy overload, requiring human experience in the later stages, which wastes manpower and resources and is inefficient. Summary of the Invention

[0004] To overcome the problems in the prior art, this invention proposes a method for correcting and predicting the transformer load rate based on heavy overload probability statistics and similar days.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] The method for correcting and predicting transformer load rates based on heavy overload probability statistics and similar days includes the following steps:

[0007] Step 1: Obtain the relevant external impact data on the prediction date, the load rates for the x days before the prediction date, and the corresponding relevant external impact data as the initial data set;

[0008] Step 2: Preprocess the initial data set to form a preprocessed data set;

[0009] Step 3: Perform feature engineering operations on the preprocessed data set to form a training data set;

[0010] Step 4: Build a load rate prediction model using LGBM, and train the load rate prediction model using the training data set to obtain a trained load rate prediction model; use the trained load rate prediction model to obtain the load prediction curve for the prediction date;

[0011] Step 5. Statistically calculate the probabilities of overload and heavy overload for each time point in the m days before the prediction date, set different types of overload probability thresholds for distribution transformation lines according to the overload probabilities, and obtain the time points with the possibility of overload;

[0012] Step 6. Obtain the load rate similar date for the prediction date;

[0013] Step 7. Correct the initial predicted load rate curve at the time points with overload probabilities according to the peak height of the load rates on the day before the prediction date and the similar date.

[0014] Further, in Step 1, the relevant external impact data includes total active power, date attribute, weather attribute, and temperature attribute.

[0015] Further, in Step 2, the preprocessing of the initial data set includes: abnormal data elimination, missing data filling, and categorical data encoding.

[0016] Further, in Step 3, the feature engineering operations specifically include:

[0017] Add date feature columns: Divide the date into 1 to 7 according to the week, add feature columns for Monday to Sunday, 1 for holidays, and 0 for other dates, and add a holiday feature column to highlight the date type;

[0018] Add the total active power features for each time point of the previous day;

[0019] Add the temperature features for each time point: Add the 96-point real-time temperature features of the local line, and use the weather forecast temperature features on the prediction date itself.

[0020] Further, in Step 6, the steps for obtaining the similar date are: Take the load rate curves for the n (n < m) days before the prediction date, find the day with the highest similarity to the load rate curve of the day before the prediction date among the n - 1 days excluding the day before the prediction date, and use the day after this date as the load rate similar date for the prediction date.

[0021] Furthermore, the method for calculating the similarity of the load factor curves is as follows:

[0022]

[0023] Among them, E i The relative error of the prediction at each time point:

[0024]

[0025] In the above formula, s represents the load rate at each time point on a similar day, t represents the load rate at each time point on the predicted day, and N represents the number of predicted time points.

[0026] Furthermore, the correction rules for the initial predicted load factor curve in step 7 include:

[0027] If the estimated probability of heavy overload at a certain point in time is 0, and the load prediction curve predicts heavy overload, then the state is corrected to normal.

[0028] If the estimated probability of heavy overload at a certain point in time is 0, and the load prediction curve shows a normal result, no correction is made.

[0029] When 0 ≤ estimated probability of heavy overload < heavy overload probability threshold β at a certain moment, the load prediction curve prediction result is not corrected;

[0030] When the estimated probability of heavy overload at a certain point in time is greater than or equal to the heavy overload probability threshold β, and the load prediction curve shows a normal state, then the larger of the actual values ​​at that point in time among similar days and the day before the prediction date is taken as the correction value.

[0031] If the estimated probability of heavy overload at a certain point in time is greater than or equal to the heavy overload probability threshold β, and the load prediction curve predicts a heavy overload, then no correction is made.

[0032] Compared with the prior art, the present invention has the following technical effects:

[0033] This invention identifies the points in time where heavy overload is likely by statistically analyzing the probability of load rate overload at various recent points in time. It focuses on correcting these points, thus improving correction efficiency. Considering the impact of short-term external factors on the peak load rate, it uses the load rate curve of the day before the prediction date to find similar days for the prediction date. Then, based on the peak load rate heights of the day before the prediction date and similar days, it corrects the initial predicted load rate curve at the points in time where heavy overload is likely, improving the prediction accuracy of the heavy overload portion and solving the problem of missed heavy overload detection. Attached Figure Description

[0034] Figure 1This is a schematic diagram illustrating the load rate correction process based on the LGBM model combined with recent overload moment probability statistics and similar days in this invention;

[0035] Figure 2 This is a statistical diagram illustrating the probability of a severe overload occurring at 96 points on a certain distribution transformer line at various times according to the present invention.

[0036] Figure 3 This is a comparison chart of the actual values ​​of the random distribution transformer line on a random day of the present invention and the prediction results of the LGBM model;

[0037] Figure 4 This is a comparison chart of the actual value, the predicted value, and the corrected predicted value of the random distribution transformer line on a random day according to the present invention. Detailed Implementation

[0038] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0039] The objective of this invention for short-term load rate correction prediction is to predict the load rate value within 24 hours of the next day, with a sampling interval of 15 minutes (96 points). It proposes a method for correcting and predicting the distribution transformer load rate based on heavy overload probability statistics and similar days, specifically including the following steps:

[0040] Step 1: Obtain relevant external impact data for the forecast date, the load rate for x days prior to the forecast date, and the corresponding relevant external impact data as the initial dataset;

[0041] The relevant external impact data includes total active power, date attribute, weather attribute, and temperature attribute.

[0042] Step 2: Preprocess the initial dataset to form a preprocessed dataset;

[0043] Since outliers can reduce the short-term predictive performance of linear regression models, data cleaning is necessary to ensure the stability of short-term predictions, reduce the interference of outliers on the model learning process, and improve prediction accuracy. Therefore, the initial dataset is preprocessed. The data preprocessing steps mainly include the following aspects:

[0044] Abnormal data removal: By setting business rules, load rates that fluctuate greatly are removed. At the same time, historical data and load rate values ​​of points near the anomaly point are used for verification. While removing outliers, outliers are also repaired.

[0045] Missing data imputation: For a small number of missing data points, the least binary search method is used to fill in the missing data. For a large number of missing data points, the measurement points are removed accordingly to reduce the outlier rate of the data and improve the availability of the data. For categorical data, null values ​​are used for imputation.

[0046] Categorical data encoding: Encode categorical data such as weather, holidays, and days of the week, converting strings into recognizable numerical classification features.

[0047] Step 3: Perform feature engineering operations on the preprocessed dataset to form a training dataset;

[0048] Load factor is influenced by various external factors such as temperature and date attributes. Therefore, in addition to the original features, other features should be considered. Not all features in the initial dataset can improve the model's predictive ability; adding some useless features can even reduce the model's predictive ability and increase training time. Selecting key features from the initial data helps improve the model's training efficiency. To address the above issues, this invention performs feature engineering on the preprocessed initial data, analyzes the feature factors affecting the magnitude and fluctuation of load factor, and selects important features to improve the model's predictive accuracy.

[0049] The main features processed in the feature engineering section of this invention are:

[0050] Add a date feature column: The electricity consumption patterns of different lines are affected by the date type. In order to strengthen the model's learning of the relationship between different types of dates and load rate, the dates are divided into 1 to 7 days of the week, and a feature column is added for Monday to Sunday. Holidays are 1, and other dates are 0. Add a holiday feature column to highlight the date type.

[0051] Add the total active power characteristics for each moment of the previous day: The real-time trend of total active power is related to the trend of load rate. Since the total active power on the prediction day is unknown, the input is the total active power of the previous day.

[0052] Add temperature features at various times: Real-time temperature changes can cause sudden increases and decreases in load. During the high temperatures of summer, users will use air conditioners intensively during the peak temperature range of the day, which will directly lead to a short-term increase in load to heavy load and overload. Therefore, we added 96 real-time temperature features of the local line and predicted the temperature features of the weather forecast for the day.

[0053] Data and features determine the upper limit of machine learning, while models and algorithms only approximate this limit. Therefore, feature engineering is of great significance for improving model training effectiveness and enhancing model performance.

[0054] Step 4: Use LGBM to build a load rate prediction model and train the load rate prediction model using the training dataset to obtain a trained load rate prediction model; use the trained load rate prediction model to obtain the load prediction curve for the prediction day.

[0055] LightGBM (Light Gradient Boosting Machine) is abbreviated as LGBM. Compared with the GBDT algorithm, LGBM has the advantages of parallel training, high training efficiency, low memory consumption, and the ability to accurately and efficiently process massive amounts of data.

[0056] The fundamental principle of LGBM (Large-Scale Decision Tree) comes from the histogram algorithm, which constructs a histogram for each feature, transforming sample traversal into histogram traversal. This process requires no additional memory, significantly reducing time complexity. LGBM employs a leaf-wise decision tree growth strategy, which traverses all leaves, locates the leaf with the largest split gain, splits it, and repeats the process. This strategy reduces errors considerably but is prone to overfitting. During training, LGBM uses a one-sided gradient algorithm, which filters out samples with small gradients, thus reducing model computation. However, LGBM is sensitive to noise, and the constructed histograms are relatively coarse, resulting in some loss of accuracy.

[0057] Predicting the load factor at 96 points can be viewed as a time series prediction problem. Machine learning is the mainstream method for solving time series problems, among which the LGBM model stands out. It can transform time series problems into supervised learning. The LGBM model can solve most time series prediction problems, supports complex data modeling, supports multivariate collaborative regression, and supports nonlinear problems.

[0058] Step 5. Statistically predict the probability of heavy overload at each time point m days before the forecast date, and set the heavy overload probability threshold for different types of distribution transformer lines based on the heavy overload probability to obtain the time points where there is a possibility of heavy overload.

[0059] The daily load rate curves of different types of distribution transformer lines vary greatly, and the peak periods of electricity load also differ. However, the peak periods of load rate for the same distribution transformer line are roughly similar each day.

[0060] The load rate of each of the 96 points for the m days preceding the forecast date is taken. Since a load rate of 80 ≤ load rate < 100 will trigger a heavy load warning, and a load rate ≥ 100 will trigger an overload warning, if it is set that a load rate greater than or equal to 78 at a certain moment, then this point may trigger a heavy overload warning, the calculation is performed by default based on the heavy overload situation. The probability of heavy overload occurring at each of the 96 points in the m days is statistically analyzed. Through statistical analysis of the heavy overload probability in the m days, a heavy overload probability threshold β is found in different distribution transformer lines. When the heavy overload probability is higher than the heavy overload probability threshold β, the risk of heavy overload at that moment is greater. The heavy overload probability threshold β can be set based on statistics and experience.

[0061] Step 6. Find days with similar load factors to the forecast date;

[0062] Since the peak height of the daily load rate is greatly affected by various recent external sudden factors, the peak height of the similar daily load rate in the recent n (n < m) days is used as a reference here.

[0063] Take the n days before the prediction day. According to the load rate curve similarity formula, find the day with the highest similarity to the load rate curve of the day before the prediction day in the n - 1 days excluding the day before the prediction day, and take the day after this date as the load rate similar day of the prediction day.

[0064] Assume that the load rate results at each time point of the similar day are s, the load rate at each time point of the prediction day is t, and the similarity is sim. The calculation formula for the load rate curve similarity at each time point is:

[0065]

[0066]

[0067] In the above formula, E i is the relative error predicted at each time point, N is the number of prediction time points. If calculated with 15 minutes as a time point, then N takes 96.

[0068] Step 7. Modify the initial predicted load rate curve at the time points with the probability of heavy overload according to the peak heights of the load rates of the day before the prediction day and the similar day.

[0069] The accuracy of the peak load rate prediction directly affects the correct rate of heavy overload determination. The short - term load rate prediction is greatly affected by various recent external sudden factors (e.g., weather, epidemic, etc.). The sudden external situations increase the uncertainty of the short - term prediction of the LGBM model, making it difficult for the LGBM model to learn the change rules in the peak load rate area. Therefore, it is necessary to increase the certainty of the short - term load rate prediction according to the recent load rate curve change rules. In this invention, the initial predicted load rate curve is modified at the time points with the probability of heavy overload according to the peak heights of the load rates of the day before the prediction day and the similar day.

[0070] Through the statistical analysis of the heavy overload probability in m days, it is found that when the heavy overload probability is higher than the heavy overload probability threshold β, the hidden danger of heavy overload at this time is greater. Therefore, the 96 points can be divided into three types, namely: heavy overload probability = 0, 0 ≤ heavy overload probability < β, heavy overload probability ≥ β.

[0071] If the probability of heavy overload at a certain time point is 0, and the LGBM prediction result is also heavy overload, then the result is corrected to a normal state (below 80). If the probability of heavy overload at a certain time point is 0, and the LGBM prediction result is also normal, then no correction is made. If at a certain time point 0 ≤ probability of heavy overload < β, then no correction is made to the LGBM prediction result. If the probability of heavy overload at a certain time point is ≥ β, and the LGBM prediction result is also normal, then the larger of the actual values ​​at that time point between similar days and the day before the prediction date is taken as the correction value. If the probability of heavy overload at a certain time point is ≥ β, and the LGBM prediction result is also heavy overload, then no correction is made.

[0072] This invention identifies the date with the highest similarity to the load rate curve of the day before the prediction date, and uses the following day as the similar day for the predicted day's load rate curve. Then, it compares the load rate at each moment between the day before the prediction date and the similar day, retaining the values ​​more likely to trigger heavy overload. Finally, it corrects the original prediction curve using a correction rule based on the probability statistics of heavy overload. Compared to the output of the LGBM model, the corrected load rate curve better addresses the possibility of missing future potential heavy overloads. This correction method reduces manual intervention and improves the accuracy of heavy overload detection.

[0073] Experimental results

[0074] The method for determining the accuracy of heavy overload at 96 points in a day in this invention experiment is as follows: Point-by-point judgment is made to determine whether the actual value and predicted value of the load rate at 96 points on the predicted day constitute a heavy overload. The overall prediction accuracy of the distribution transformer line in a day is PR. total The calculation formula is:

[0075]

[0076] Where, N correct To predict the correct number of points, N total Total points (N) total =96).

[0077] To verify the feasibility and usability of this invention, numerous experiments were conducted, and some of the experimental results are as follows:

[0078] (I) Statistical analysis of the probability of overload occurrence at different times on different lines recently

[0079] The probability results of heavy overload occurring at various time points in the recent 96 points of a distribution transformer line are as follows: Figure 2As shown, the time periods for heavy overloads are approximately from 0:00 to 4:30 and from 18:30 to 23:45 each day. The probability of heavy overloads occurring at night is much higher than in the early morning. The probability of heavy overloads occurring during the period from 4:30 to 18:30 is zero. Therefore, in determining whether a distribution transformer has experienced a heavy overload, it is only necessary to adjust the time points where there is a potential for heavy overload.

[0080] (II) Comparison of LGBM model predictions with actual values

[0081] A line was randomly selected, and dates were randomly filtered to predict the load rate for June 24, 2022. The results are as follows. Figure 3 As shown, the dashed line represents the initial load rate prediction curve output by the LGBM model, and the solid line represents the actual load rate curve on the prediction day. It's clear that the curves output by the LGBM model alone are quite similar in overall shape, especially in the middle section where the load rate is lower. However, the load rates are higher at the early morning and nighttime periods, reaching severe overload at certain points. The LGBM model's predictions for these periods are much lower than the actual results, failing to provide timely and accurate warnings of potential severe overload risks during operation.

[0082] (III) Comparison of revised prediction results with LGBM model prediction results and actual values

[0083] like Figure 4 As shown, the dashed line represents the load rate prediction curve after correction by the method of this invention, the solid line represents the actual load rate curve, and the dotted line represents the initial prediction curve output by LGBM. It can be seen that the corrected curve is closer to the actual curve in the parts with higher load rates at both ends, which can detect potential heavy overload hazards in time.

[0084] The accuracy of the overload correction prediction results for different distribution transformer lines is compared with the initial prediction results in Tables 1 and 2.

[0085] Table 1. Comparison of Overload Correction Prediction Results for Different Distribution Transformer Lines on June 26, 2022

[0086]

[0087] Table 2 Comparison of Overload Correction Prediction Results for Different Distribution Transformer Lines on July 24, 2022

[0088]

[0089] Tables 1 and 2 compare the predictions and corrections made using the LGBM model for June 26, 2022, and July 24, 2022, respectively, on 10 randomly selected lines of different types. The study found that the average prediction accuracy for transformer overload was improved across the 10 lines of different types. Furthermore, this correction method is more intelligent and automated compared to manual correction based on subjective experience.

[0090] The experimental results above show that the short-term transformer load rate correction prediction method based on heavy overload probability statistics and similar days can effectively improve the accuracy of short-term transformer heavy overload prediction.

[0091] Since the probability of each distribution line being under heavy overload at any of the 96 times of the day is relatively low, the large difference between the proportion of heavy overload and non-heavy overload leads to a relatively high accuracy of the overall load rate curve, but a relatively low accuracy in predicting whether a heavy overload condition has occurred. Furthermore, in practice, the LGBM model performs poorly in predicting heavy overload conditions.

[0092] Therefore, this invention proposes a method for short-term load rate correction prediction. It statistically analyzes the probability of heavy load overload at various recent time points, identifies the time points where heavy overload is likely, and focuses on correcting the time points where heavy overload is likely, thereby improving correction efficiency. Considering the impact of short-term external factors on the peak load rate, it uses the load rate curve of the day before the prediction date to find the load rate similar to the prediction date. Then, based on the peak load rate height of the day before the prediction date and the similar day, it corrects the initial predicted load rate curve at the time points where heavy overload is likely, thereby improving the prediction accuracy of the heavy overload portion and solving the problem of missed heavy overload detection.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting and predicting transformer load rate based on heavy overload probability statistics and similar days, characterized in that, It includes the following steps: Step 1: Obtain the relevant external influence data on the prediction date, the load rates in the x days before the prediction date, and the corresponding relevant external influence data as the initial data set; Step 2: Preprocess the initial data set to form a preprocessed data set; Step 3: Perform feature engineering operations on the preprocessed data set to form a training data set; Step 4: Use LGBM to construct a load rate prediction model, and train the load rate prediction model with the training data set to obtain a trained load rate prediction model; use the trained load rate prediction model to obtain the load prediction curve on the prediction date; Step 5. Statistically calculate the probabilities of overload and heavy overload at each time point in the m days before the prediction date, and set the overload probability thresholds for different types of distribution transformer lines according to the heavy overload probabilities to obtain the time points with the possibility of heavy overload; Step 6. Obtain the load rate similar day of the prediction date, including: take the load rate curves in the n days before the prediction date, n < m, find the day with the highest similarity to the load rate curve of the day before the prediction date in the n - 1 days excluding the day before the prediction date, and take the day after this date as the load rate similar day of the prediction date; Step 7. Correct the initial predicted load rate curve at the time points with heavy overload probabilities according to the peak heights of the load rates on the day before the prediction date and the similar day, including: When the estimated heavy overload probability at a certain time point = 0, and the load prediction curve predicts heavy overload, it is corrected to the normal state; When the estimated heavy overload probability at a certain time point = 0, and the load prediction curve predicts normal, no correction is made; When 0 ≤ the estimated heavy overload probability < the heavy overload probability threshold β at a certain time point, the prediction result of the load prediction curve is not corrected; When the estimated heavy overload probability at a certain time point ≥ the heavy overload probability threshold β, and the load prediction curve predicts the normal state, take the larger true value of this time point in the similar day and the day before the prediction date as the correction value; When the estimated heavy overload probability at a certain time point ≥ the heavy overload probability threshold β, and the load prediction curve predicts heavy overload, no correction is made.

2. The method for correcting and predicting transformer load rate based on heavy overload probability statistics and similar days as described in claim 1, characterized in that, The relevant external influence data in Step 1 includes total active power, date attribute, weather attribute, and temperature attribute.

3. The method for correcting and predicting transformer load rate based on heavy overload probability statistics and similar days as described in claim 1, characterized in that, The preprocessing of the initial data set in Step 2 includes: abnormal data elimination, missing data filling, and categorical data encoding.

4. The method for correcting and predicting transformer load rate based on heavy overload probability statistics and similar days as described in claim 1, characterized in that, The feature engineering operations in Step 3 specifically include: Adding a date feature column: Divide the date into 1 to 7 by week, add feature columns for Monday to Sunday, with holidays being 1 and other dates being 0, and add a holiday feature column to highlight the date type; Adding the total active power features at each time point of the previous day; Adding the temperature features at each time point: Add the 96-point real-time temperature features of the local line, and use the weather forecast temperature feature on the prediction date itself.

5. The method for correcting and predicting transformer load rate based on heavy overload probability statistics and similar days as described in claim 1, characterized in that, In Step 6, to find the day with the highest similarity to the load rate curve of the day before the prediction date in the n - 1 days excluding the day before the prediction date, the load rate curve similarity calculation is adopted, specifically: in, The relative error of the prediction at each time point: In the above formula, The load factor results are for similar days at different times. To predict the load rate at each time point of the day, N is the number of time points to predict.

Images