Power load abnormal value prediction method

Abstract

The invention relates to a power load abnormal value prediction method, and the method comprises the steps: inputting data into a prediction model, wherein the data comprises meteorological data, holiday and festival data of a to-be-predicted region in a period, and the power load data of each industry; by introducing meteorological data and holiday and festival data, the influence of third-party factors on the prediction result can be reduced to a certain extent, and the prediction accuracy is improved. Meanwhile, the power load normal value data and the power load abnormal value data are respectively trained through different models, the trained models are fitted to obtain the prediction model, and the final prediction model can combine the advantages of different models, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The present application relates to the technical field of power load prediction, and in particular to a method for predicting abnormal power load values. Background Art

[0002] With the rapid development of the economy and society, people's demand for electricity is increasing day by day, and the volatility of power load values ​​is also increasing. The power system is a top priority in the field of infrastructure, playing a key supporting and promoting role in economic and social development. The fluctuation of power load affects the construction of power infrastructure. Therefore, accurate prediction of abnormal values ​​of power load is crucial to the entire power system.

[0003] At present, the stability of the power system has been greatly improved compared with before, but with the continuous increase of the global population and the rapid development of the economy, it is still necessary to use power resources more efficiently. Power load data is closely related to power demand, and the supply and demand relationship of the power system should maintain a dynamic balance, so that power companies can formulate correct operation plans and strategies to adapt to the power market demand.

[0004] At present, the detection of power load anomalies mainly includes model-based methods, machine learning methods, and statistical methods. Power load forecasting mainly adopts time series forecasting, traditional machine learning forecasting, neural network forecasting, and expert system forecasting. Time series forecasting predicts future loads by analyzing historical power loads. It uses statistical methods and time series analysis techniques to establish a mathematical model to predict future power demand and load. Machine learning forecasting uses support vector machines, decision trees, random forests and other algorithms to analyze and model a large amount of collected data to obtain predicted loads. Expert system forecasting uses artificial intelligence technology and domain expert knowledge to build an expert system, and makes forecasts based on specific needs by analyzing historical data. Neural networks model power demand into a neural network and train it. By learning from existing data, future power demand can be inferred.

[0005] Although the existing power load anomaly prediction methods can help power companies predict future power demand and load, optimize energy scheduling and planning, and improve the efficiency and reliability of power systems, the existing power load anomaly prediction methods still have the problem of unsatisfactory overall prediction results. Summary of the invention

[0006] In order to at least to some extent overcome the problem that the overall prediction results of the power load abnormal value prediction method in the related art are not ideal, the present application provides a power load abnormal value prediction method.

[0007] The scheme of this application is as follows:

[0008] A method for predicting abnormal values ​​of power load, comprising:

[0009] Collect meteorological data, holiday data and electricity load data of various industries in the forecasted area within the period;

[0010] Input the collected data into the pre-trained prediction model to obtain the prediction results of the abnormal value of power load in the next cycle in the area to be predicted;

[0011] Among them, training the prediction model includes:

[0012] The normal value data of the power load and the abnormal value data of the power load are trained respectively by different models, and the trained models are fitted to obtain the prediction model.

[0013] Preferably, the method further comprises:

[0014] Obtain historical meteorological data, holiday data, and electricity load data of various industries in the area to be predicted as sample data;

[0015] interpolating the sample data;

[0016] Performing power load abnormal value detection on the interpolated sample data, and dividing the interpolated sample data into power load normal value data and power load abnormal value data;

[0017] The power load abnormal value data is trained by the first model and the second model, the prediction results of the first model and the second model are passed as features to the third model, and the training of the power load abnormal value data is completed by the third model;

[0018] Training the power load normal value data through the fourth model;

[0019] The trained third model and the fourth model are fitted to obtain the prediction model.

[0020] Preferably, interpolating the sample data comprises:

[0021] Establishing a regression model to interpolate missing values ​​in the sample data;

[0022] According to the interpolation result of the regression model, the accuracy of the regression model is measured using the mean square error;

[0023] Adjust the parameters of the regression model according to the mean square error measurement result until the accuracy of the regression model meets the requirements, thereby obtaining an interpolation model;

[0024] The missing values ​​in the sample data are interpolated using the interpolation model.

[0025] Preferably, performing power load abnormal value detection on the interpolated sample data includes:

[0026] Divide the imputed sample data into quartile ranges;

[0027] Determine the outlier interval based on the divided quartile range;

[0028] The power load outlier data is identified in the interpolated sample data according to the determined outlier interval.

[0029] Preferably, the training of power load abnormal value data includes:

[0030] Construct a two-layer XGBoost algorithm model as the first model;

[0031] Analyze the periodic characteristic quantity of the abnormal value data of the power load through the reference layer, and output the reference layer prediction result;

[0032] By comparing the predicted results of the benchmark layer with the measured values ​​through the real-time layer, the dynamic error is obtained;

[0033] The time characteristics are iteratively calculated based on the dynamic error and the baseline layer prediction results are added to output the real-time layer prediction results.

[0034] Preferably, the training of power load abnormal value data further includes:

[0035] Perform a stationarity test on the sequence data in the power load anomaly data to observe whether the sequence is stationary, and perform differential processing on the non-stationary sequence data;

[0036] The self-correlation coefficient and partial autocorrelation coefficient of the sequence data after stationary processing are obtained respectively, and the optimal order is obtained by analyzing the autocorrelation graph and partial autocorrelation graph;

[0037] Construct an ARIMA model as the second model according to the optimal order;

[0038] The performance of the second model is verified. After the performance verification is passed, the power load abnormal value data is input into the second model to obtain the prediction result of the second model.

[0039] Preferably, the training of power load abnormal value data further includes:

[0040] Construct the LightGBM algorithm model as the third model, and perform data fitting on the prediction results of the first model and the second model through time series cross-validation; including:

[0041] The prediction results of the first model and the second model are used as a data set;

[0042] In chronological order, the data set is gradually expanded, and the data set is divided into training set and validation set according to the time point;

[0043] The entire data set is used as a validation set through iteration to complete the training of the third model.

[0044] Preferably, the training of normal value data of power load includes:

[0045] Normalize or standardize the normal value data of power load;

[0046] Determine the kernel function and its parameters;

[0047] Constructing a support vector machine model as a fourth model according to the processed power load normal value data, and determining hyperparameters of the fourth model; the hyperparameters are used to control the fault tolerance and fitting accuracy of the fourth model;

[0048] Determine the hyperplane and support vector of the fourth model that maximizes the interval by an optimization algorithm, and complete the training of the fourth model;

[0049] The trained fourth model is used to predict the normal value data of the power load, and the prediction result is obtained through the mapping of the kernel function.

[0050] Preferably, fitting the trained third model and the fourth model to obtain the prediction model includes:

[0051] The trained third model and the fourth model are fitted by linear interpolation, and the average value of the prediction results of the third model and the fourth model at the same time is used as the prediction result after fitting;

[0052] The prediction accuracy of the prediction results after fitting is analyzed through residual analysis.

[0053] Preferably, before training the power load abnormal value data, the method further comprises: extracting features from the power load abnormal value data using Ridge regression.

[0054] The technical solution provided by this application may have the following beneficial effects:

[0055] The power load abnormal value prediction method in the present application includes: collecting meteorological data, holiday data and power load data of various industries in the area to be predicted within a period; inputting the collected data into a pre-trained prediction model to obtain the power load abnormal value prediction result of the area to be predicted in the next period; wherein, training the prediction model includes: respectively training the power load normal value data and the power load abnormal value data through different models, fitting the trained models, and obtaining the prediction model.

[0056] The data used to input the prediction model in this technical solution include the meteorological data, holiday data and power load data of various industries in the area to be predicted within the period. By introducing meteorological data and holiday data, the influence of third-party factors on the prediction results can be reduced to a certain extent, and the prediction accuracy can be improved. At the same time, in this application, different models are used to train the normal value data of power load and the abnormal value data of power load, and the trained models are fitted to obtain the prediction model. The final prediction model can combine the advantages of different models to improve the prediction accuracy.

[0057] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0059] Figure 1 It is a flowchart of a method for predicting abnormal values ​​of power load provided by an embodiment of the present application;

[0060] Figure 2 It is a flowchart of training a prediction model in a method for predicting abnormal values ​​of power load provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0062] Figure 1 is a flow chart of a method for predicting abnormal values ​​of power load provided by an embodiment of the present application, referring to Figure 1 , a method for predicting abnormal values ​​of power load, comprising:

[0063] S11: Collect meteorological data, holiday data and power load data of various industries in the forecasted area within the period;

[0064] It should be noted that the cycle in this embodiment can be set to one month, and the meteorological data, holiday data and power load data of various industries within one month can well reflect whether there are abnormal values ​​in the power load of the area to be predicted.

[0065] Meteorological data is a set of data that reflects the weather. Different weather conditions have a certain impact on power load. For example, the probability of abnormal power load values ​​on sunny days is relatively low, while the probability of abnormal power load values ​​on rainy days is relatively high. Therefore, in this embodiment, weather is included in the prediction of abnormal power load values ​​to improve the prediction accuracy.

[0066] The impact of holidays on power load is illustrated below:

[0067] 1. For spring and autumn, the probability of abnormal power load values ​​is relatively high in summer and winter due to the high power demand of air conditioning and heating equipment;

[0068] 2. For enterprises, the probability of abnormal power load values ​​on weekdays is relatively higher than on weekends.

[0069] Therefore, in this embodiment, incorporating holidays into the prediction of abnormal power load values ​​can improve the prediction accuracy.

[0070] S12: input the collected data into a pre-trained prediction model to obtain a prediction result of abnormal power load values ​​in the next cycle in the area to be predicted;

[0071] Among them, training the prediction model includes:

[0072] The normal value data of power load and the abnormal value data of power load are trained respectively through different models, and the trained models are fitted to obtain the prediction model.

[0073] For specific training prediction model process, please refer to Figure 2 ,include:

[0074] S21: Obtain historical meteorological data, holiday data, and power load data of various industries in the area to be predicted as sample data;

[0075] S22: interpolate sample data;

[0076] S23: Performing power load abnormal value detection on the interpolated sample data, and dividing the interpolated sample data into power load normal value data and power load abnormal value data;

[0077] S24: training the power load abnormal value data through the first model and the second model, passing the prediction results of the first model and the second model as features to the third model, and completing the training of the power load abnormal value data through the third model;

[0078] S25: training the power load normal value data through the fourth model;

[0079] S26: Fitting the trained third model and the fourth model to obtain a prediction model.

[0080] Specifically, the sample data is interpolated, including:

[0081] Build a regression model to interpolate missing values ​​in sample data;

[0082] According to the interpolation results of the regression model, the mean square error is used to measure the accuracy of the regression model;

[0083] According to the mean square error measurement result, the parameters of the regression model are adjusted until the accuracy of the regression model meets the requirements, and the interpolation model is obtained;

[0084] The missing values ​​in the sample data are imputed using the imputation model.

[0085] The regression model is as follows:

[0086] Y=β0+β1X1+β2X2+…+β p X p +ε#(1)

[0087] Among them, Y is the power load data, X1, X2…X p is time, β0, β1, …β p are model parameters and ε is the error term.

[0088] The formula for calculating the mean square error is:

[0089]

[0090] Where n is the total number of observations, yi is the i-th observation value, is the interpolated value of the ith observation.

[0091] It should be noted that the power load abnormal value detection is performed on the interpolated sample data, including:

[0092] Divide the imputed sample data into quartile ranges;

[0093] Determine the outlier interval based on the divided quartile range;

[0094] The power load outlier data is identified in the interpolated sample data according to the determined outlier interval.

[0095] The interquartile range method for detecting outliers in power load includes the following steps:

[0096] First quartile (Q1): 25% of the data points in the data set are less than or equal to this value;

[0097] Second quartile (Q2): 50% of the data points in the data set are less than or equal to this value;

[0098] Third quartile (Q3): 75% of the data points in the data set are less than or equal to this value;

[0099] IQR is the difference between Q3 and Q1, that is:

[0100] IQR=Q3-Q1#(3)

[0101] Set the outlier threshold, and define the outlier interval as data points below the outlier threshold Q1-1.5×IQR or above the outlier threshold Q3+1.5×IQR. In this way, outliers are identified and processed in the next step.

[0102] It should be noted that in this embodiment, the power load anomaly data is trained through a hybrid model, and the hybrid model includes: a two-layer XGBoost algorithm, an ARIMA algorithm, and a LightGBM algorithm. Among them, the two-layer XGBoost algorithm model uses supervised learning to reduce the amount of training for predicting outlier detection in different industries, the ARIMA algorithm model processes time series data to capture the trend and seasonal characteristics of the data, and the LightGBM algorithm model is used as a meta-model to improve the prediction accuracy by using its gradient boosting framework and tree-based learning algorithm.

[0103] Specifically, the training of power load anomaly data includes:

[0104] Construct a two-layer XGBoost algorithm model as the first model;

[0105] Analyze the periodic characteristic quantity of the abnormal value data of the power load through the reference layer, and output the reference layer prediction result;

[0106] By comparing the predicted results of the benchmark layer with the measured values ​​through the real-time layer, the dynamic error is obtained;

[0107] The time characteristics are iteratively calculated based on the dynamic error and the baseline layer prediction results are added to output the real-time layer prediction results.

[0108] The definition function of the XGBoost algorithm is:

[0109]

[0110] In the formula, is the predicted value; x i is the i-th sample; k is the model of the k-th tree; K is the total number of trees; H is the function space composed of all decision trees.

[0111] The loss function of XGBoost is:

[0112]

[0113] In the formula, is the measured value yi and the predicted value ; Ω(f) is the regularization function used to control the complexity of the model; v is the number of leaves; ω is the weight value of the leaf node.

[0114] In the two-layer collaborative structure, the baseline layer will first give the load forecast results, and the real-time layer will correct the forecast results according to the deviation of the baseline layer data, and establish a forecast model that is closest to the actual working conditions based on the actual situation, thereby making the forecast results more reliable.

[0115] The training of power load anomaly data also includes:

[0116] Perform a stationarity test on the sequence data in the power load anomaly data to observe whether the sequence is stationary, and perform differential processing on the non-stationary sequence data;

[0117] The self-correlation coefficient and partial autocorrelation coefficient of the sequence data after stationary processing are obtained respectively, and the optimal order is obtained by analyzing the autocorrelation graph and partial autocorrelation graph;

[0118] Construct an ARIMA model as the second model according to the optimal order;

[0119] The performance of the second model is verified. After the performance verification is passed, the power load abnormal value data is input into the second model to obtain the prediction result of the second model.

[0120] The ARIMA model includes the autoregressive (AR) process, the moving regression (MA) process, and the difference (I) process.

[0121] 1) The difference I process calculates the difference between the time series at time t and t-1, and makes the non-stationary series stationary, which is defined as:

[0122] Δy(x)=y(x+1)-y(x), (x=0,1,2,…)#(8)

[0123] 2) AR is an autoregressive process, and p is an autoregressive term, which is used to describe the relationship between the current value and the historical value. It is defined as:

[0124]

[0125] Among them, y t is the current value, μ is the constant term, p is the order, γ i is the autocorrelation coefficient, ∈ t For error.

[0126] 3) MA is a moving regression process, q is the number of moving average terms, and this process focuses on the accumulation of error terms in the autoregressive process. It is defined as:

[0127]

[0128] Moving average can effectively eliminate random fluctuations in predictions.

[0129] The steps to construct the ARIMA algorithm model are:

[0130] 1) Perform a stationary test on the sequence data to observe whether the sequence is stationary, and perform differential processing on the non-stationary sequence to make it stationary;

[0131] 2) The ACF and PACF of the sequence data after stationary processing are obtained respectively, and the optimal orders p and q are obtained by analyzing the autocorrelation graph and partial autocorrelation graph;

[0132] ACF formula:

[0133]

[0134] PACF formula:

[0135]

[0136] 3) Obtain the p, q, and d values ​​from the above steps, obtain the ARIMA model, verify the performance of the model, and output the predicted value.

[0137] The training of power load anomaly data also includes:

[0138] Construct the LightGBM algorithm model as the third model, and perform data fitting on the prediction results of the first model and the second model through time series cross-validation; including:

[0139] The prediction results of the first model and the second model are used as a data set;

[0140] In chronological order, the data set is gradually expanded, and the data set is divided into training set and validation set according to the time point;

[0141] The entire data set is used as a validation set through iteration to complete the training of the third model.

[0142] The LightGBM algorithm model gradually improves the prediction accuracy of the model by continuously optimizing the residual of the model in an iterative manner. The expression of the LightGBM algorithm model is:

[0143]

[0144] in, Represents the predicted output value of LightGBM, f t (d i) represents the predicted value of the t-th regression tree for the XGBoost algorithm prediction time series results and the ARIMA prediction time series results as the i-th sample of the input data set, and T represents the total number of regression trees;

[0145] The objective function of the LightGBM algorithm model is composed of the loss function and the regularization term. The expression of the objective function is:

[0146] obj t =L t +Ω t #(14)

[0147] The expressions of the loss function and regular penalty term are as follows:

[0148]

[0149] Among them, obj t is the objective function of the t-th tree, L t and Ω t are the loss function and regular penalty term of the t-th tree, J is the number of leaf nodes, γ is the penalty coefficient of leaf nodes, M is the total number of samples, Represents the predicted value of the t-1th tree for the i-th sample.

[0150] During the training process, the time series cross-validation method will be used to fit the data. The specific steps are as follows:

[0151] (1) Expand the data set gradually in chronological order. The previous data is used as the training set, and the later data is used as the validation set.

[0152] (2) In step 1, the training set is the first t1 time points, and the validation set is t1+1 to t2;

[0153] (3) In step 2, the training set is extended to the first t2 time points, and the validation set is t2+1 ​​to t3;

[0154] (4) And so on, until all time points are used as validation sets.

[0155] Finally, the hybrid model composed of the first model, the second model and the third model is trained and verified until a prediction result that meets the preset requirements is output.

[0156] It should be noted that the prediction results of XGBoost and ARIMA are passed to the LightGBM model as input features and trained uniformly in the LightGBM model.

[0157] The XGBoost model implements the patching of outliers, and ARIMA implements the normal and reasonable data trend of the data set. The data results of the two models are time series results, but in different directions. Training in the LightGBM model can combine the characteristics of the two to obtain patch values ​​that conform to the time series trend.

[0158] The two models used cross-validation method during training to prevent overfitting.

[0159] The prediction results of XGBoost and ARIMA are the prediction results of time series data such as historical meteorological data, historical electricity load data of various industries, holiday information, etc. They belong to the same dimension during the training process of the LightGBM model and do not need to be fitted.

[0160] It should be noted that the training of normal value data of power load includes:

[0161] Normalize or standardize the normal value data of power load;

[0162] Determine the kernel function and its parameters;

[0163] A support vector machine model is constructed as a fourth model according to the processed power load normal value data, and a hyperparameter of the fourth model is determined; the hyperparameter is used to control the fault tolerance and fitting accuracy of the fourth model;

[0164] Determine the hyperplane and support vector of the fourth model that maximizes the interval through the optimization algorithm, and complete the training of the fourth model;

[0165] The trained fourth model is used to predict the normal value data of the power load, and the prediction result is obtained through the mapping of the kernel function.

[0166] The specific training steps of the support vector machine model are as follows:

[0167] 1. Data preprocessing: normalize or standardize the data.

[0168] 2. Kernel function selection: Select an appropriate kernel function and determine its parameters.

[0169] 3. Model building: To build the SVR model using training data, it is necessary to determine the hyperparameters C and ε, which control the model's fault tolerance and fitting accuracy respectively.

[0170] 4. Model training: By solving the optimization problem, the hyperplane and support vector that maximize the margin are found. This can be achieved using an optimization algorithm (such as the sequential minimum optimization algorithm).

[0171] 5. Prediction: Use the trained model to predict the data, map it to high-dimensional space through the kernel function, determine the predicted value, and obtain the prediction result;

[0172] It should be noted that the trained third model and the fourth model are fitted to obtain a prediction model, including:

[0173] The trained third model and the fourth model are fitted by linear interpolation, and the average value of the prediction results of the third model and the fourth model at the same time is used as the prediction result after fitting;

[0174] The prediction accuracy of the prediction results after fitting is analyzed through residual analysis.

[0175] It should be noted that since the prediction results of the third and fourth models are time series predictions, the linear interpolation method is used for fitting. For the prediction results of the two models at the same time, the average value is used as the prediction result after fitting. Finally, the residual analysis is used to evaluate the fitting effect and prediction accuracy of the model.

[0176] Model fit test:

[0177] (1) Residual analysis

[0178]

[0179] Among them, yt is the observed value, is the model prediction value.

[0180] (2) Residual distribution test

[0181] On the basis of residual analysis, the residual distribution is evaluated by drawing a residual histogram to test whether the residual distribution satisfies the normal distribution. This step is very important for correlation testing.

[0182] Residual autocorrelation test

[0183] Check whether the model adequately captures the time series characteristics of the data, and check whether the residuals show time series autocorrelation based on the residual distribution test.

[0184] Model prediction accuracy evaluation:

[0185] The root mean square error is used to test the accuracy of the model prediction. Based on the residual analysis, the residual sum of squares is calculated as follows:

[0186]

[0187] Where S is the residual sum of squares and εt is the residual.

[0188] The formula for calculating the root mean square error is:

[0189]

[0190] Among them, the smaller the root mean square error value, the smaller the difference between the predicted value and the observed value, which means that the prediction accuracy of the model is higher; the larger the root mean square error value, the greater the difference between the predicted value and the observed value, which means that the prediction accuracy of the model is lower.

[0191] It should be noted that, before training the power load anomaly data, the method further includes: extracting features from the power load anomaly data using Ridge regression.

[0192] Since there are many variables in the prediction process, Ridge regression is used in this embodiment to extract features from the power load abnormal value data. The specific formula is:

[0193]

[0194] Among them, β is the vector parameter of the model, y i is the target variable for the ith observation, X i is the feature vector of the ith observation, n is the number of samples, p is the number of features, and λ is the regularization parameter used to control the strength of regularization.

[0195] In summary, the data used to input the prediction model in this technical solution include the meteorological data, holiday data and power load data of various industries in the area to be predicted within the period. By introducing meteorological data and holiday data, the influence of third-party factors on the prediction results can be reduced to a certain extent, and the prediction accuracy can be improved. At the same time, in this application, different models are used to train the normal value data of power load and the abnormal value data of power load, respectively, and the trained models are fitted to obtain the prediction model. The final prediction model can combine the advantages of different models to improve the prediction accuracy.

[0196] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0197] It should be noted that, in the description of this application, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of this application, unless otherwise specified, the meaning of "plurality" refers to at least two.

[0198] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0199] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0200] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0201] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0202] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0203] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0204] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for predicting abnormal values ​​of power load, characterized in that: include: Collect meteorological data, holiday data and electricity load data of various industries in the forecasted area within the period; Input the collected data into the pre-trained prediction model to obtain the prediction results of the abnormal value of power load in the next cycle in the area to be predicted; Among them, training the prediction model includes: The normal value data of the power load and the abnormal value data of the power load are trained respectively by different models, and the trained models are fitted to obtain the prediction model.

2. The method according to claim 1, characterized in that The method further comprises: Obtain historical meteorological data, holiday data, and electricity load data of various industries in the area to be predicted as sample data; interpolating the sample data; Performing power load abnormal value detection on the interpolated sample data, and dividing the interpolated sample data into power load normal value data and power load abnormal value data; The power load abnormal value data is trained by the first model and the second model, the prediction results of the first model and the second model are passed as features to the third model, and the training of the power load abnormal value data is completed by the third model; Training the power load normal value data through the fourth model; The trained third model and the fourth model are fitted to obtain the prediction model.

3. The method according to claim 2, characterized in that Interpolating the sample data includes: Establishing a regression model to interpolate missing values ​​in the sample data; According to the interpolation result of the regression model, the accuracy of the regression model is measured using the mean square error; Adjust the parameters of the regression model according to the mean square error measurement result until the accuracy of the regression model meets the requirements, thereby obtaining an interpolation model; The missing values ​​in the sample data are interpolated using the interpolation model.

4. The method according to claim 2, characterized in that: The power load anomaly detection is performed on the interpolated sample data, including: Divide the imputed sample data into quartile ranges; Determine the outlier interval based on the divided quartile range; The power load outlier data is identified in the interpolated sample data according to the determined outlier interval.

5. The method according to claim 2, characterized in that: The training of power load anomaly data includes: Construct a two-layer XGBoost algorithm model as the first model; Analyze the periodic characteristic quantity of the abnormal value data of the power load through the reference layer, and output the reference layer prediction result; By comparing the predicted results of the benchmark layer with the measured values ​​through the real-time layer, the dynamic error is obtained; The time characteristics are iteratively calculated based on the dynamic error and the baseline layer prediction results are added to output the real-time layer prediction results.

6. The method according to claim 2, characterized in that The training of power load anomaly data also includes: Perform a stationarity test on the sequence data in the power load anomaly data to observe whether the sequence is stationary, and perform differential processing on the non-stationary sequence data; The self-correlation coefficient and partial autocorrelation coefficient of the sequence data after stationary processing are obtained respectively, and the optimal order is obtained by analyzing the autocorrelation graph and partial autocorrelation graph; Construct an ARIMA model as the second model according to the optimal order; The performance of the second model is verified. After the performance verification is passed, the power load abnormal value data is input into the second model to obtain the prediction result of the second model.

7. The method according to claim 2, characterized in that The training of power load anomaly data also includes: Construct the LightGBM algorithm model as the third model, and perform data fitting on the prediction results of the first model and the second model through time series cross-validation; including: The prediction results of the first model and the second model are used as a data set; In chronological order, the data set is gradually expanded, and the data set is divided into training set and validation set according to the time point; The entire data set is used as a validation set through iteration to complete the training of the third model.

8. The method according to claim 2, characterized in that: The training of normal value data of power load includes: Normalize or standardize the normal value data of power load; Determine the kernel function and its parameters; Constructing a support vector machine model as a fourth model according to the processed power load normal value data, and determining hyperparameters of the fourth model; the hyperparameters are used to control the fault tolerance and fitting accuracy of the fourth model; Determine the hyperplane and support vector of the fourth model that maximizes the interval by an optimization algorithm, and complete the training of the fourth model; The trained fourth model is used to predict the normal value data of the power load, and the prediction result is obtained through the mapping of the kernel function.

9. The method according to claim 2, characterized in that: Fitting the trained third model and the fourth model to obtain the prediction model includes: The trained third model and the fourth model are fitted by linear interpolation, and the average value of the prediction results of the third model and the fourth model at the same time is used as the prediction result after fitting; The prediction accuracy of the prediction results after fitting is analyzed through residual analysis.

10. The method according to claim 2, characterized in that Before training the power load abnormal value data, the method further includes: using Ridge regression to extract features from the power load abnormal value data.

Images