Load forecasting and early warning method and system for transformer area containing distributed resources

Abstract

The invention belongs to the technical field of power distribution networks, and discloses a load prediction and early warning method and system for a transformer area containing distributed resources, and the method comprises the steps: combining a two-dimensional time sequence data sample set of a high-risk transformer area with a transformer area feature operation data set according to the transformer area and a timestamp, and generating a prediction model training sample data set; building a lightweight gradient boosting tree as a main prediction model, inputting a training sample data set for training, and optimizing hyper-parameters of the main prediction model by adopting a Bayesian optimization algorithm; and establishing a residual error correction model based on local weighted Gaussian process regression, superposing a load prediction result of a prediction day of the main prediction model of the to-be-predicted transformer area with a residual error correction value of a prediction day of the residual error correction model to obtain a final load prediction result, and outputting transformer area weight / overload early warning information. According to the method, the LGBM is adopted as the main prediction model for load prediction, the residual error correction model is adopted for residual error correction, and the robustness and adaptability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution networks, and more specifically, relates to a load prediction and early warning method and system for a station area containing distributed resources. Background Art

[0002] Overloading or overloading distribution transformers (DTs) is a major cause of equipment failure and power outages, posing a serious threat to the safe and stable operation of distribution networks. Specifically, during traditional high-load periods (such as peak summer and winter electricity demand), DTs are prone to prolonged overloading or overloading, which can easily lead to equipment damage and power outages. Furthermore, the recent influx of new distributed resources, such as photovoltaics and charging stations, into medium and low voltage distribution networks has significantly increased load volatility and uncertainty, exacerbating the risk of overloading or overloading DTs in both the forward and reverse directions, posing new challenges to the safe operation of distribution networks.

[0003] Carrying out distribution transformer load forecasting and heavy / overload warnings helps operations and maintenance personnel assess the transformer's carrying capacity in advance and take preventive measures, which is an important means of ensuring regional power supply security. Current research mainly includes two types of methods: one is a prediction model based on machine learning, which can comprehensively consider multi-dimensional data such as users and electricity usage patterns at a smaller model scale, but the simulation results are less accurate; the other is a hybrid model of single-stage / multi-stage time series prediction that integrates deep learning such as neural networks. Although these models have strong nonlinear fitting capabilities, they are not well adapted to small sample scenarios such as newly built substations, have limited model interpretability, require long training times, are not updated in a timely manner, and are susceptible to time series distribution drift. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention aims to provide a load forecasting and early warning method and system for distributed resource substations with a small model scale and high prediction accuracy, so as to solve the problems that the existing methods are difficult to balance lightweight models with high accuracy and high debugging costs.

[0005] To achieve the above objectives, according to one aspect of the present invention, a load forecasting and early warning method for a station area containing distributed resources is provided, comprising: S1: Screen forward and reverse heavy / overload high-risk areas as target areas; S2: Obtain the target area load data, meteorological data, and time data to generate a two-dimensional time series data sample set; S3: Obtain the target area's ledger data, user attribute data, and electricity consumption data, and process them to generate characteristic operation data sets for each area; S4: Merge the two-dimensional time series data sample set and the substation feature operation data set by substation and timestamp to generate a prediction model training sample data set; S5: Build a Light Gradient Boosting Machine (LGBM) as the main prediction model, input the training sample dataset for training, and use the Bayesian optimization algorithm to optimize the main prediction model hyperparameters; S6: According to steps S2 to S4, relevant data of all the substations to be predicted from the prediction date to the month before are obtained, the data set to be predicted is formed and input into the trained main prediction model to obtain the load forecast results of the main prediction model from the prediction date to the month before; S7: Process the data obtained in step S6, calculate and filter out the 7 days with the weather most similar to the day to be predicted, and construct a residual data set based on the meteorological data of these 7 days, the load forecast results of the main forecast model, and the forecast residual data; S8: A residual correction model is established based on Local Weighted Gaussian Process Regression (LGPR). The residual dataset is used as the training set, and the load forecast results of the main forecast model and meteorological data on the forecast day are used as the forecast input to obtain the load residual correction value on the forecast day. S9: Superimpose the load forecast result of the main forecast model on the forecast day of the substation to be forecasted with the load residual correction value on the forecast day to obtain the final load forecast result; based on the predicted load level, evaluate the operation risk of the distribution transformer and issue a heavy / overload warning information for the substation.

[0006] Preferably, in step S1, the process of screening forward and reverse heavy / overload high-risk areas is: S101: Enter the forecast period and select the substations with forward and reverse heavy / overload occurrences in the past three months and the same month of the previous year from the historical database; S102: Search for substations with an absolute load factor exceeding 60% within one week before the forecast period and within ±1 week of the same period of the previous year; S103: If it is a holiday, additional search is performed for the substations whose absolute load rate exceeds 60% within ±1 week of the same holiday in the previous year.

[0007] Preferably, the specific process in step S2 is: S201: Obtain the target area's load data and meteorological data. The load data is the area's historical daily load, covering the three months prior to the forecast month and the same month last year. Meteorological data includes temperature, humidity, and rainfall. Time data is generated from the timestamp of the load data, including month, date, day of the week, whether it is a holiday, etc. The accuracy of all time series data is hourly, with a total of 24 data points per day for each area. S202: Eliminate abnormal load data that exceeds twice the capacity of the corresponding target substation, and use the least squares regression method and linear interpolation method to fill in the missing load data and time data; S203: Calculate the quartile values and the maximum and minimum load values for the processed target area load data on a monthly and area-by-area basis, mark abnormal load data, and obtain the abnormality rate of the target area data on a daily basis; S204: using a boxplot method to eliminate abnormal points in meteorological data such as temperature and humidity, and using a least squares regression method to fill in missing meteorological data; S205: Integrate the processed and cleaned data to generate a two-dimensional time series data sample set.

[0008] Preferably, in step S203, the load data in the target area is processed by the box plot method, and the specific steps of obtaining the data abnormality rate in the target area on a daily basis are as follows: (1) Detect abnormal load data according to the degree to which the data deviates from the average level, and calculate the upper quartile Q3 and the lower quartile Q1 of the sample. The interquartile range is recorded as IQR = Q3-Q1, and the interval is The data other than the above data are marked as abnormal data; (2) the proportion of abnormal data in the three days before the previous day to the total data volume in the three days is calculated day by day as the abnormal rate.

[0009] Preferably, the specific steps of step S3 are: S301: Obtain the target area's ledger data, user daily electricity consumption, and user attribute data, where the user daily electricity consumption data covers the three months before the forecast month and the same month last year; S302: For each day, count the number of active users and power consumption of different types of users in the target area and within the target area; an active user is defined as one whose power consumption exceeds a threshold on two days in the last three days; S303: Calculate the total power consumption of the target area and each type of user in the target area every day; S304: Integrate the daily active capacity and power consumption of various users and combine them with user attribute information to form a comprehensive characteristic data set of the substation operation.

[0010] Preferably, the specific process of step S5 is: S501: Build a lightweight gradient boosting tree as the main prediction model; S502: The hourly load of each substation in the sample data set is used as the prediction target value, and the remaining features are used as input values to train the main prediction model; S503: Using the Bayesian optimization algorithm to determine the optimal hyperparameters of the main prediction model.

[0011] Furthermore, in step S6, the substations to be predicted include but are not limited to the high-risk substations screened in step S1, and the relevant data obtained include all data types involved in steps S2 and S3, and the data time range covers the prediction day and the month before it.

[0012] Furthermore, in step S7, for the area to be predicted, based on meteorological data such as daily maximum temperature, average temperature, minimum temperature, daily rainfall, average humidity and weather conditions (such as sunny, cloudy, etc.), the Euclidean distance between the predicted day and each day in the previous month is first calculated, and the time attenuation factor is generated by combining the difference in the number of days between the two days. The product of distance and time attenuation factor is used as the weather similarity index. The first seven days with the highest weather similarity to the predicted area are selected, and the residual data of five dimensions, namely temperature, humidity, rainfall, and the residual between the load prediction results of the main prediction model and the actual observation values, are constructed to form a residual dataset.

[0013] Preferably, the specific process of step S8 is: S801: Using the temperature, humidity, rainfall, and load prediction results of the main prediction model in the residual data set as features, and the residual between the load prediction results of the main prediction model and the actual observation value as the target, as a training set; S802: The predicted temperature, humidity, rainfall, and load prediction results of the main prediction model on the prediction day are used as a test set; S803: Build a local weighted Gaussian process regression model, input the training set and the test set into the residual correction model and train it to generate the load residual correction result of the predicted day of the substation to be predicted.

[0014] Furthermore, in step S9, the area where the predicted absolute value of the load reaches 80% of the rated capacity of the distribution transformer and lasts for more than 2 hours is determined to be a heavy-loaded area, and the area where the predicted absolute value of the load reaches 100% of the distribution transformer and lasts for more than 2 hours is determined to be an overloaded area, and the rest are normal areas. Taking the 1h level load data in the present invention as an example, the area operation risk determination is shown in the following table:

[0015] The present invention also discloses a load forecasting and warning system for distributed resource areas, including a memory and a processor. The memory stores a computer program, and when the computer program is run by the processor, it executes the steps of the load forecasting and warning method for distributed resource areas.

[0016] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art: (1) The innovative introduction of multi-dimensional substation operation characteristic data such as the number of active users and the amount of active photovoltaic users’ electricity consumption significantly improved the generalization ability of the prediction model. It can effectively simulate the impact of various types of distributed resource access on the substation load and accurately depict the load change trends of different types of substations. (2) Using various data cleaning and missing value filling methods such as box plot method and least squares regression method, we effectively reduced the interference of outliers and missing data on model training results, ensuring data quality and model robustness; (3) Using LGBM as the main prediction model for load forecasting has the advantages of small model size, low training cost, and strong interpretability. It can effectively improve the efficiency of model training and iteration and reduce the drift problem of time series prediction.

[0017] (4) Taking the forecast date as the benchmark, the seven days of data with the highest weather similarity to the target date within the one-month time window before the forecast date were selected to form a training set. An improved local weighted Gaussian process regression model was established, and a rolling residual correction scheme was implemented on the output results of the LGBM main forecast model. This scheme selected seven days of highly correlated data within the past month as training samples, which not only significantly reduced the consumption of computing resources, but also avoided the weather fluctuation interference that might be introduced by using only short-term continuous data, thereby improving the robustness and adaptability of the model.

[0018] (5) By introducing an improved local weighting mechanism, the limitation of traditional Gaussian process regression methods, such as those that require multiple days of new distribution data to be accumulated in the training set before adapting to changes and compensating for residuals, is effectively overcome. This method can adapt to and track the distribution changes of substation load data more quickly, thereby significantly improving the accuracy of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the method of the present invention.

[0020] Figure 2 This is a load forecast result diagram of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0022] The present invention proposes a load forecasting and early warning method and system for substations containing distributed resources, which can take into account both lightweight models and high accuracy, effectively solving the problems of high debugging costs and poor applicability of existing methods, and has good prediction effects on distribution transformers in various substations containing distributed resources.

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown, the load forecasting and early warning method for a distributed resource area in this embodiment includes the following steps: S1: Screen forward and reverse heavy / overload high-risk areas as target areas; S2: Obtain the target area load data, meteorological data, and time data to generate a two-dimensional time series data sample set; S3: Obtain the target area's ledger data, user attribute data, and electricity consumption data, and process them to generate characteristic operation data sets for each area; S4: Merge the two-dimensional time series data sample set and the substation feature operation data set by substation and timestamp to generate a prediction model training sample data set; S5: Build a Light Gradient Boosting Machine (LGBM) as the main prediction model, input the training sample dataset for training, and use the Bayesian optimization algorithm to optimize the main prediction model hyperparameters; S6: According to steps S2 to S4, relevant data of all the substations to be predicted from the prediction date to the month before are obtained, the data set to be predicted is formed and input into the trained main prediction model to obtain the load forecast results of the main prediction model from the prediction date to the month before; S7: Process the data obtained in step S6, calculate and filter out the 7 days with the weather most similar to the day to be predicted, and construct a residual data set based on the meteorological data of these 7 days, the load forecast results of the main forecast model, and the forecast residual data; S8: A residual correction model is established based on Local Weighted Gaussian Process Regression (LGPR). The residual dataset is used as the training set, and the load forecast results of the main forecast model and meteorological data on the forecast day are used as the forecast input to obtain the load residual correction value on the forecast day. S9: Superimpose the load forecast result of the main forecast model on the forecast day of the substation to be forecasted with the load residual correction value on the forecast day to obtain the final load forecast result; based on the predicted load level, evaluate the operation risk of the distribution transformer and issue a heavy / overload warning information for the substation.

[0025] Step S1: Screen high-risk areas as target areas. The process is as follows: S101: Enter the forecast period and select the substations in the province that have experienced forward and reverse heavy / overload in the past three months and the same month of the previous year from the historical database; S102: Search for substations with an absolute load factor exceeding 60% within one week before the forecast period and within ±1 week of the same period of the previous year; S103: If it is a holiday, additional search is performed for the substations whose absolute load rate exceeds 60% within ±1 week of the same holiday in the previous year.

[0026] Specifically, in step S103, since the corresponding Gregorian calendar dates of traditional holidays (such as Spring Festival, etc.) are different each year, in order to avoid data deviation caused by different holiday times, the same holiday in the previous year and the week before and after are supplemented in this step.

[0027] The specific process of step S2 includes the following steps: S201: Obtain the target area's load data and meteorological data. The load data is the area's historical daily load, covering the three months prior to the forecast month and the same period last year. The meteorological data includes temperature, humidity, and rainfall. The time data is generated from the timestamp of the load data, including the hour, date, month, day of the week, whether it is a holiday, etc. The accuracy of all time series data is at the hourly level, with a total of 24 data points per day for each area. In this step, the historical daily load of the substation is the time series data of the three-phase average power load. The collection interval of the original load data is 15 minutes, that is, data is collected every 15 minutes from 00:00 to 23:59 every day. Since the collection interval of meteorological data is 1 hour, in order to facilitate data alignment, the collection frequency of load data is downsampled to once per hour, so that each substation obtains 24 load data points every day. The load data is recorded as , is the load data at time t, t={1,2,…,24}, and the rated capacity of the distribution transformer is denoted as p cap .

[0028] Specifically, in step S201, the generation process of time data is as follows: based on the timestamp of the load data, the corresponding time features such as the month, day of the month, and hour of the day are extracted; at the same time, according to the week information, Monday to Friday are marked as working days (marked as 0), and Saturday, Sunday and statutory holidays are marked as rest days (marked as 1).

[0029] S202: Eliminate abnormal historical load data that exceeds twice the capacity of the corresponding target substation, and use the least squares regression method and linear interpolation method to fill in the missing load data and time data; S203: Calculate the quartiles and the maximum and minimum load values for the processed target area load data on a monthly and area-by-area basis. Mark abnormal load data and obtain the abnormality rate of the target area data on a daily basis. In this step, firstly, based on the monthly load data of each target area, the upper quartile (Q3), lower quartile (Q1) and maximum and minimum load values of the sample are counted, and the interquartile range (IQR) is calculated. Data points outside of the above range are marked as abnormal load data. Subsequently, the ratio of abnormal load data in the three days before the target area to the total data volume in the three days is counted on a daily basis, which is used as the abnormal rate of the target area on that day and is used for subsequent abnormal data removal and data quality control.

[0030] S204: using a boxplot method to eliminate abnormal points in meteorological data such as temperature and humidity, and using a least squares regression method to fill in missing meteorological data; S205: Integrate the processed and cleaned data to generate a two-dimensional time series data sample set; Specifically, in step S205, the characteristics of the formed two-dimensional time series data sample set are: each row corresponds to the load, weather, time and other related data of a certain substation at a certain moment.

[0031] Through the above-mentioned data cleaning, data anomalies and missing phenomena caused by abnormal collection and noise interference of measurement equipment can be effectively solved, the quality of time series data in the distribution transformer area can be improved, and a good data basis can be provided for distribution transformer load prediction and heavy / overload warning.

[0032] The specific process of step S3 includes the following steps: S301: Obtain the target area's ledger data, user daily electricity consumption, and user attribute data. The user's daily electricity consumption data covers the three months before the forecast month and the same month last year. User attribute tags include: residential user ID, photovoltaic power generation facility ID, charging pile user ID, etc. The ledger data includes the city to which the area belongs, whether it is an urban area, and the area capacity. S302: For each day, count the number of active users and electricity consumption for the target area and different types of users within the target area (such as residents, photovoltaic users, charging piles, etc.). An active user is defined as one whose electricity consumption (generation) exceeds the threshold on two days in the past three days. The total active electricity consumption is defined as the average electricity consumption (generation) of this type of active user in the past three days; Specifically, in step S302, the criteria for determining active users are as follows: for residential users, the daily electricity consumption exceeds 2 kWh; for photovoltaic and charging pile users, the daily electricity consumption / generation exceeds 5 kWh. The relevant criteria can be adjusted accordingly based on the model prediction results and the prediction period.

[0033] S303: Obtain the total power consumption of the target area and various types of users (such as residents, photovoltaics, charging piles, etc.) in the target area every day.

[0034] S304: Integrate the daily active capacity and power consumption data of various users, and combine them with user attributes to form a comprehensive characteristic operation data set for the substation.

[0035] Specifically, the characteristics of the comprehensive characteristic operation data set of the substation formed in step S304 are: each row corresponds to the total electricity consumption, active electricity consumption, active residents' electricity consumption, active photovoltaic power generation and other data of a substation on a certain day.

[0036] Through the above steps, processing and generating comprehensive characteristic operation data of the substation can greatly enhance the generalization ability of the model and effectively support the comprehensive modeling and accurate simulation of the load characteristics of various substations containing distributed resources (such as photovoltaics, charging piles, etc.).

[0037] The specific process of step S5 includes the following steps: S501: Build a lightweight gradient boosting tree (LGBM) model as the main prediction model. The hyperparameters of the lightweight gradient boosting tree include tree structure parameters (maximum leaf tree, maximum tree depth, minimum number of leaf node samples), regularization parameters (data sampling ratio, sampling frequency, selected feature ratio for each iteration), and learning rate. S502: The hourly load of each substation in the sample data set is used as the prediction target value, and the remaining features are used as input values to train the lightweight gradient boosting tree; S503: Using the Bayesian optimization algorithm to determine the optimal hyperparameters of the lightweight gradient boosting tree.

[0038] Specifically, in step S502, the sample data set is randomly divided into a training sample set and a validation sample set, where the training sample set accounts for 80% and the validation sample set accounts for 20%. The main prediction model is updated with data and retrained every two weeks.

[0039] Specifically, the principle of lightweight gradient boosting tree is as follows: The Lightweight Gradient Boosting Machine (LGBM) is an efficient gradient boosted decision tree (GBDT) framework. This method iteratively trains multiple classification and regression trees, with each new tree fitting the residuals of the previous model's predictions. The negative gradient of the loss function (i.e., the residual) guides the growth of the new tree, gradually bringing the model's predicted values closer to the true values. The Lightweight Gradient Boosting Machine (LGBM) uses an efficient histogram binning method to accelerate feature splitting and introduces a leaf-prioritized splitting strategy (unlike traditional layer-by-layer splitting), further improving the model's training speed and prediction accuracy. The Lightweight Gradient Boosting Machine is shown below: ; in, represents the predicted value of the u-th sample, represents the u-th sample, is the set space of trees, represents the model of the nth tree, is the number of models.

[0040] The lightweight gradient boosting tree uses a leaf-based tree splitting strategy. In each round, only the leaf node with the largest splitting gain is selected, rather than parallel splitting of nodes at the same level. The gain function can be expressed as: ; ; in, is the model gain, is the splitting gain, leaf is the leaf, and current bestleaf indicates the leaf node to be split with the largest gain in the current iteration round in the tree splitting strategy generated by leaves; 、 Represent the gradients of the left and right leaf nodes respectively, 、 where λ represents the gradient and Hessian value accumulation of the left and right leaf nodes, respectively, and λ is the L2 regularization coefficient. Unlike the traditional GBTD algorithm, LGBM does not sort continuous features and traverse the split points when calculating the gain. Instead, it first histograms the features (bins them) and searches only for the optimal split point on the bounds of the finite bins, significantly accelerating the algorithm and reducing space complexity.

[0041] In order to evaluate the prediction effect, the prediction result evaluation index is established. The mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) are used to measure the deviation between the predicted value and the actual value. RMSE is used as the performance index of LGBM to correct the prediction error of the main prediction model.

[0042] Specifically, in step S6, the substations to be predicted include but are not limited to the high-risk substations selected in step S1. The relevant test data obtained includes all data types involved in steps S2 and S3, and the data time range covers the prediction date and the month before it. This test data is processed according to the methods of steps S2 to S4 and input into the main prediction model trained in step S5 to obtain the load forecast results of the main prediction model.

[0043] Specifically, in step S7, the data obtained in S6 is processed to calculate and select the seven days with the weather most similar to the day to be predicted. The specific implementation steps are: (1) Based on the normalized meteorological data of the forecast day and each day in the previous month (including the daily maximum temperature T max , average temperature T avg , minimum temperature T min , daily rainfall R, average humidity H and weather conditions W, etc.), where weather conditions W are coded as 0, 1, 2, etc. according to the degree of similarity. First, calculate the Euclidean distance D between the predicted day and each day in the previous month i : Where, subscript 0 represents the forecast day, i represents the historical day, γ is the weather state weight, is the predicted maximum daily temperature, The highest temperature on record. To predict the average daily temperature, is the historical average temperature on day i, is the predicted daily minimum temperature, The lowest temperature on the day in history, To predict daily rainfall, is the historical daily rainfall, To predict the average daily humidity, is the historical average humidity on day i, To predict the weather conditions for the day, For historical daily weather conditions.

[0044] (2) Next, combine the difference in days between the two dates , calculate the time decay factor ,in is the adjustment coefficient, is the historical date, For the predicted date, it can be taken as 0.1 in this embodiment.

[0045] (3) Finally, the product of the Euclidean distance and the time decay factor is used as the weather similarity index S i : , weather similarity index S i The smaller it is, the more similar the meteorological conditions on the historical day are to those on the forecast day.

[0046] The specific process of step S8 includes the following steps: S801: Using the temperature, humidity, rainfall, and load prediction results of the main prediction model in the residual data set as features, and the residual between the load prediction results of the main prediction model and the actual observation value as the target, as a training set; S802: The predicted temperature, humidity, rainfall, and load prediction results of the main prediction model on the prediction day are used as a test set; S803: Build a local weighted Gaussian process regression model, input the training set and the test set into the residual correction model and train it to generate the load residual correction result of the predicted day of the substation to be predicted.

[0047] Specifically, the Local Weighted Gaussian Process Regression (LWGPR) model is trained with rolling data daily. The principle of LWGPR is as follows: Gaussian process regression (GPR) is a nonparametric regression method based on kernel functions and Bayesian theory that can generate probability distribution predictions for any input point. Locally weighted Gaussian process regression (LWGPR) builds on GPR by introducing feature similarity weights. This method assigns a dynamic weight to each sample based on the distance between the prediction point and historical samples in feature space, thereby improving the model's responsiveness to local feature changes and achieving more accurate regression predictions. The specific implementation process is as follows: (1) Prepare the training set described in S801 ,in, is the u-th training sample, including the normalized LGBM prediction value, temperature, humidity, and rainfall at that time point, is the output corresponding to the u-th training sample, that is, the residual between the load forecast result of the main forecast model and the actual observation value at that moment. N is the total number of samples.

[0048] (2) Select a test sample from the test set , For the feature at a certain time point, the similarity weight is calculated for each sample in the training set. For the u-th sample in the training set, the formula is: ; in, is the sample feature and The similarity weight of is the bandwidth parameter, Represents the Euclidean distance between two features.

[0049] (3) Determine the expansion multiple of each sample in the training set based on the similarity weight: ; in, is the maximum expansion multiple, are the sample features in the training set and The maximum similarity weight of For the jth training sample, expand the training set U according to the relevant results to obtain the local training set: .

[0050] (4) In the local training set Construct a Gaussian process using the RBF kernel function: ; in, For two input feature vectors F and The kernel function value between is the signal variance, and l is the length scale.

[0051] (5) Test samples Perform Gaussian process regression prediction to obtain the predicted mean : ; in, Output vector for the local training set, is the kernel matrix of the expanded local training set , is the kernel matrix of the test sample and the local sample, for The transpose of . for and The kernel function value between For test samples and The kernel function value between is the noise variance, is the identity matrix, Used to add noise terms to the kernel matrix K.

[0052] (6) Repeat steps (2) to (5) for each time point feature in the test set to obtain the residual correction values of all the time points to be predicted.

[0053] The following provides specific implementation cases to further explain the technical solutions of the present invention: The load of 1,000 substations in a province on September 1, 2024 is predicted. The specific prediction process is as follows: Step 1) Obtain a list of 6,314 high-risk substations across the province with heavy / overload events from June to August 2024 and September 2023, as well as those with load factors exceeding 60% between August 25 and 31, 2024, and between August 25 and September 8, 2023. Simultaneously obtain historical load data and relevant meteorological data for these substations from September 2023 and June to August 2024. Remove and mark outliers from the aggregated load and meteorological data, and use least squares regression and linear interpolation to infill missing and anomalous data.

[0054] Step 2) Obtain daily user usage data, user attribute data, and district ledger data for high-risk substations in September 2023 and June-August 2024, and process them to generate a daily feature operation dataset for each substation. Merge this with the dataset from Step 1) based on the timestamp to obtain a training sample dataset for the prediction model. The specific data format is as follows:

[0055] Step 3) Build and train the LGBM using the sample data from 6,314 substations before September 1, 2024, as the training set. Use the Bayesian optimization algorithm to adjust and optimize the hyperparameters of the LGBM.

[0056] Step 4) Based on Steps 1) and 2), extract the relevant input data for the 1,000 substations to be forecasted between August 1 and September 1, 2024. The load on September 1 is the target, and the temperature, humidity, and other meteorological information are the forecast values obtained one day in advance. The test dataset is input into the trained LGBM primary forecasting model to obtain preliminary load forecasts for the 1,000 substations from August 1 to September 1. Furthermore, the residuals between the primary forecasting model's predicted values and the actual load from August 1 to 31 are calculated.

[0057] Step 5) Based on the meteorological characteristics of September 1, such as daily maximum temperature, average temperature, minimum temperature, daily rainfall, average humidity, and weather conditions (such as sunny, cloudy, etc.), the Euclidean distance is calculated with the corresponding meteorological data of each day in the previous month, and the time decay factor is generated by combining the difference in days between the two dates. The product of the Euclidean distance and the time decay factor is used as the weather similarity index, and the 7 days with the smallest index (i.e., the 7 days most similar to September 1) are selected as reference.

[0058] Step 6) For a specific forecast area, using the seven days with the most similar weather conditions selected in Step 5) as samples, extract the model-predicted values, temperature, humidity, and rainfall for these seven days as training features, and the residuals as training targets to construct a residual training set. Simultaneously, the model-predicted values, temperature, humidity, and rainfall for the same area on September 1st serve as the test feature set. Based on this, an LWGPR residual correction model is constructed. The residual training set and test feature set are input into the model to obtain the corrected residual for the load for that area on September 1st.

[0059] Step 7) Add the LGBM main prediction model's predicted value for the substation to be predicted on September 1st to the LWGPR residual to obtain the final load forecast for that substation. Repeat steps 5)-6) to obtain load forecast values for all substations to be predicted. Based on the load forecast results, if three consecutive load forecast values in a substation exceed 80% (100%) of the substation capacity, a heavy load (overload) warning is issued. Three models, LGBM alone, LGBM + original GPR, and LGBM + LWGPR, were selected for comparative testing. The load forecast results for the 1,000 substations to be predicted are shown in the table below.

[0060]

[0061] As shown in the table, the LGBM+LWGPR forecasting method performs best in terms of average MAE, average RMSE, and average MAPE. Compared with the LGBM model alone, the proposed method improves the average MAE and average RMSE by 30.4% and 19.4%, respectively, significantly improving the overall accuracy of load forecasting. Compared with the LGBM+GPR method, the average MAE and average RMSE also improve by 20.9% and 8.0%, respectively.

[0062] The prediction results of a certain area containing distributed photovoltaics are as follows: Figure 2 As shown in the figure, the LGBM model first predicts the general trend of load changes in the substation area. Compared with the GPR model, the LWGPR model better corrects the prediction deviation at the load peak. This improvement is due to the local weighted Gaussian process regression method's ability to fully exploit historical samples close to the prediction point, achieving a "larger near, smaller far" weight distribution, further enhancing the model's ability to fit local features and correct residual errors.

[0063] Another embodiment of the present invention provides a load forecasting and warning system for a station area containing distributed resources, including a memory and a processor, wherein a computer program is stored on the memory, and when the computer program is run by the processor, the steps of the above-mentioned load forecasting and warning method for a station area containing distributed resources are executed.

[0064] In summary, this paper proposes a load forecasting and heavy / overload warning method and system for substations with distributed resources. This system employs a short-term load forecasting strategy that combines a lightweight gradient boosting tree (LGBM) main prediction model with a locally weighted Gaussian process regression (LWGPR) residual correction model. This method balances model lightweightness with high accuracy, effectively addressing the high commissioning costs and poor applicability of existing methods. It demonstrates excellent forecasting results for distribution transformers in various substations with distributed resources.

[0065] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A load forecasting and early warning method for a distributed resource area, characterized by: include: S1: Screen forward and reverse heavy / overload high-risk areas as target areas; S2: Obtain the target area load data, meteorological data, and time data to generate a two-dimensional time series data sample set; S3: Obtain the target area's ledger data, user attribute data, and electricity consumption data, and process them to generate characteristic operation data sets for each area; S4: Merge the two-dimensional time series data sample set and the substation feature operation data set by substation and timestamp to generate a prediction model training sample data set; S5: Build a lightweight gradient boosting tree as the main prediction model, input the training sample data set for training, and use the Bayesian optimization algorithm to optimize the hyperparameters of the main prediction model; S6: According to steps S2 to S4, relevant data of all the substations to be predicted from the prediction date to the month before are obtained, a sample data set to be predicted is formed and input into the trained main prediction model to obtain the load forecast results of the main prediction model from the prediction date to the month before; S7: Process the data obtained in step S6, calculate and filter out the 7 days with the weather most similar to the day to be predicted, and construct a residual data set based on the meteorological data of these 7 days, the load forecast results of the main forecast model, and the forecast residual data; S8: A residual correction model is established based on local weighted Gaussian process regression, with the residual data set as the training set and the load forecast results of the main forecast model and meteorological data on the forecast day as the forecast input, to obtain the load residual correction value on the forecast day; S9: Superimpose the load forecast result of the main forecast model on the forecast day of the substation to be forecasted with the load residual correction value on the forecast day to obtain the final load forecast result; based on the predicted load level, evaluate the operation risk of the distribution transformer and issue a heavy / overload warning information for the substation.

2. A load forecasting and early warning method for distributed resource areas according to claim 1, characterized in that the steps In S1, the process of screening forward and reverse heavy / overload high-risk areas is as follows: S101: Enter the forecast period and select the substations with forward and reverse heavy / overload occurrences in the past three months and the same period of the previous year from the historical database; S102: Search for substations with an absolute load factor exceeding 60% within one week before the forecast period and within ±1 week of the same period of the previous year; S103: If it is a holiday, additional search is performed for the substations whose absolute load rate exceeds 60% within ±1 week of the same holiday in the previous year.

3. The load forecasting and early warning method for a distributed resource area according to claim 1 is characterized in that: The specific process in step S2 is: S201: Obtaining target area load data and meteorological data; S202: Eliminate abnormal load data that exceeds twice the capacity of the corresponding target substation, and use the least squares regression method and linear interpolation method to fill in the missing load data and time data; S203: Calculate the quartile values and the maximum and minimum load values for the processed target area load data on a monthly and area-by-area basis, mark abnormal load data, and obtain the abnormality rate of the target area data on a daily basis; S204: using a boxplot method to eliminate abnormal points in the meteorological data, and using a least squares regression method to fill in the missing meteorological data; S205: Integrate the processed and cleaned data to generate a two-dimensional time series data sample set.

4. The load forecasting and early warning method for a distributed resource area according to claim 1 is characterized in that: The specific steps of step S3 are: S301: Obtain the target area's ledger data, user daily electricity consumption, and user attribute data, where the user daily electricity consumption data covers the three months before the forecast month and the same month last year; S302: For each day, count the number of active users and power consumption of different types of users in the target area and within the target area; an active user is defined as one whose power consumption exceeds a threshold on two days in the last three days; S303: Calculate the total power consumption of the target area and each type of user in the target area every day; S304: Integrate the daily active capacity and power consumption of various users and combine them with user attribute information to form a comprehensive characteristic data set of the substation operation.

5. The load forecasting and early warning method for a distributed resource area according to claim 1 is characterized in that: The specific process of step S5 is: S501: Build a lightweight gradient boosting tree as the main prediction model; S502: The hourly load of each substation in the sample data set is used as the prediction target value, and the remaining features are used as input values to train the main prediction model; S503: Using the Bayesian optimization algorithm to determine the optimal hyperparameters of the main prediction model.

6. The load forecasting and early warning method for distributed resource areas according to claim 1 is characterized in that In step S6, the substations to be predicted include but are not limited to the high-risk substations screened in step S1, and the relevant data obtained include all data types involved in steps S2 and S3, and the data time range covers the prediction day and the month before it.

7. The load forecasting and early warning method for distributed resource areas according to claim 1 is characterized in that: In step S7, for the area to be predicted, based on the meteorological data, the Euclidean distance between the predicted day and each day in the previous month is calculated, and the time attenuation factor is generated by combining the difference in days between the two days. , the product of distance and time decay factor is used as the weather similarity index; The first seven days with the highest similarity to the weather in the substation to be predicted are selected, and the residual data of five dimensions, including temperature, humidity, rainfall, load prediction results of the main prediction model, and the residual between the load prediction results of the main prediction model and the actual observation value, are constructed to construct a residual data set.

8. The load forecasting and early warning method for a distributed resource area according to claim 1 is characterized in that: The specific process of step S8 is: S801: Using the temperature, humidity, rainfall, and load prediction results of the main prediction model in the residual data set as features, and the residual between the load prediction results of the main prediction model and the actual observation value as the target, as a training set; S802: The predicted temperature, humidity, rainfall, and load prediction results of the main prediction model on the prediction day are used as a test set; S803: Build a local weighted Gaussian process regression model, input the training set and the test set into the residual correction model and train it to generate the load residual correction result of the predicted day of the substation to be predicted.

9. The load forecasting and early warning method for a distributed resource area according to claim 1 is characterized in that: In step S9, the substation where the predicted absolute value of the load reaches 80% of the rated capacity of the distribution transformer and lasts for more than 2 hours is determined as a heavy-loaded substation, and the substation where the predicted absolute value of the load reaches 100% of the distribution transformer and lasts for more than 2 hours is determined as an overloaded substation. The rest are normal substations.

10. A load forecasting and warning system for distributed resource areas, comprising a memory and a processor, characterized in that: The memory stores a computer program, which, when executed by the processor, executes the steps of the load prediction and early warning method for a distributed resource area according to any one of claims 1 to 9.

Images