A triangular fuzzy prediction method for power load based on MEEMD and optimal combination integration

Abstract

This invention relates to a triangular fuzzy prediction method for power load based on MEEMD and optimal combination integration, which overcomes the information loss caused by the fact that power load prediction values ​​are usually real-value sequences compared to existing technologies. The invention includes the following steps: acquiring the dataset; decomposing the dataset; performing single-model prediction; constructing a triangular fuzzy optimal combination prediction model; and obtaining the triangular fuzzy prediction results for power load. This invention performs empirical mode decomposition on the raw power load data, constructs triangular fuzzy numbers, and uses three models—XGBoost, MSVR, and MLP—for combined prediction. It utilizes the MEEMD algorithm and optimal combination integration technology to improve prediction accuracy and mitigate prediction risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, in particular to a power load triangular fuzzy prediction method based on MEEMD and optimal combination integration. BACKGROUND

[0002] With the development of smart grid technology, a large number of renewable energy sources are connected to the power system, and the demand for electricity is also increasing. Because of the characteristics of electricity storage difficulty, its production and consumption must be synchronized. If the power generation cannot meet the demand, it will cause power grid failure, and if the supply exceeds the demand, it will cause waste of energy and resources. Therefore, accurate power load prediction plays an important role in maintaining power supply and demand balance, reducing power generation cost and ensuring safe and stable operation of the power system.

[0003] In actual situations, there are many factors affecting the change of power load, and these factors often have different degrees of uncertainty. Therefore, using fuzzy mathematical theory to predict power load is more reasonable and closer to the actual situation. For example, describing the power load data of a day, the left endpoint and the right endpoint of the triangular fuzzy number can represent the minimum power load and the maximum power load of the day, respectively, and the midpoint of the triangular fuzzy number can represent the average power load of the day. Compared with real value time series, triangular fuzzy number is an effective form of describing the instantaneous information contained in power load.

[0004] In the method of short-term power load prediction, the traditional time series analysis method is poor in processing nonlinear data, and the multivariate ensemble empirical mode decomposition (MEEMD) is an improved adaptive data analysis method that can obtain effective information of power load data. By adding auxiliary white noise to reduce the influence of modal aliasing, the time series is decomposed into several stationary intrinsic mode functions and a residual sequence, and then the decomposed sequences are predicted respectively, and finally integrated, which has the effect of "divide and conquer". However, due to the different information sources or the bias of the function form of the model setting, the single prediction model may produce large prediction error, thereby affecting the accuracy of decision-making, so the single prediction model has certain defects.

[0005] Therefore, it is necessary to design a power load triangular fuzzy prediction method based on MEEMD and optimal combination integration, which is applied to the field of power prediction by performing empirical mode decomposition and optimal combination integration on the original power load triangular fuzzy data. SUMMARY

[0006] The purpose of the present application is to solve the defect of information loss caused by the fact that the power load prediction value in the prior art is usually a real value sequence, and to provide a power load triangular fuzzy prediction method based on MEEMD and optimal combination integration to solve the above problems.

[0007] To achieve the above object, the technical scheme of the present application is as follows:

[0008] The power load triangular fuzzy prediction method based on MEEMD and optimal combination integration comprises the following steps:

[0009] 11) Data set acquisition: Obtain the hourly power load data in units of days, and calculate the maximum value R, the minimum value L and the average value M of the power load of each day respectively;

[0010] 12) Data set decomposition: The MEEMD method is used to decompose the power load data to obtain 11 groups (11*3) of intrinsic mode functions that eliminate the mixing phenomenon;

[0011] 13) Single model prediction: The first k data are used as the training set, and the last n-k data are used as the test set. XGBoost, MSVR and MLP are used to predict the power load data of the next n-k days for each group of intrinsic mode functions;

[0012] 14) Construction of triangular fuzzy optimal combination prediction model: The prediction results of the three single prediction models are used to construct a triangular fuzzy optimal combination prediction model for each group of intrinsic mode functions, which minimizes the sum of squared errors of power load prediction;

[0013] 15) Obtaining of power load triangular fuzzy prediction results: The triangular fuzzy optimal combination prediction model of each group of intrinsic mode functions is solved by software to obtain the prediction results of the 11 groups of intrinsic mode functions, and then linear superposition is performed to obtain the integrated prediction results, i.e. the power load prediction results.

[0014] The data set decomposition comprises the following steps:

[0015] 21) Divide the original data by its own standard deviation to eliminate the influence of order of magnitude and dimension, and further prepare for subsequent superposition of white noise;

[0016] 22) Attach normal distribution white noise with a standard deviation of NTSD and a mean of 0 to the original data to generate a new set of data containing white noise;

[0017] 23) Perform empirical mode decomposition on the data containing white noise to obtain 11 groups of intrinsic mode functions;

[0018] 24) Set the additional calculation number as Ne, repeat steps 23) and 24) for Ne times, and finally obtain 11 groups of intrinsic mode function sets containing white noise;

[0019] 25) Sum each set of eigenmode functions containing white noise and divide by Ne+1 to obtain the mean function of the eigenmode functions containing white noise in the set, and multiply by the standard deviation to restore the order of magnitude, and then obtain 11 sets of eigenmode functions that eliminate the mixing phenomenon.

[0020] The performing single model prediction comprises the following steps:

[0021] 31) Set the XGBoost model using the sklearn library function of python, and the parameters are set as follows: the integration learning rate is set to 0.1, the number of weak evaluators in the integration is set to 300, the maximum depth of each tree is set to 5, the sum of the second derivatives on the leaf nodes is set to 1 similar to the sample weight, the sample ratio when randomly sampling is set to 0.8, the ratio of randomly sampled features when generating a tree is set to 0.8, the L2 regularization term is used, and the parameter lambda of the regularization term is set to 1;

[0022] 32) Set the MSVR model using the Gaussian kernel function rbf, and the parameters are set as follows: the parameter sigma affecting the radial action range of rbf is set to 560000000, the penalty coefficient C is set to 580000, the error precision for stopping training is set to 1e-10, Epsilon is set to 1.04, and the epsilon-tube is specified, the epsilon-tube contains the points predicted by the distance less than Epsilon from the actual value in the training loss function;

[0023] 33) Set the MLP model using the sklearn library function of python, and the parameters are set as follows: the relu activation function is used, the number of input layer neurons is 3, the number of hidden layer neurons is set to 130, and the number of output layer neurons is 3;

[0024] 34) The power load data set of the previous k days is respectively brought into the XGBoost model, the MSVR model, and the MLP model for training, and then the power load data set of the next n-k days is respectively brought into the above three models for single model prediction, to obtain the prediction results of the above three models.

[0025] The constructing triangular fuzzy optimal combination prediction model comprises the following steps:

[0026] 41) Define the prediction errors of the three endpoints L, M, and R in the triangular fuzzy number, which reflect the prediction effect of the minimum value, the mean value, and the maximum value, and set the errors of the three endpoints to be equally important to obtain the prediction error of the triangular fuzzy number;

[0027] 42) Define the triangular fuzzy optimal combination optimization objective function and its constraint condition as follows:

[0028]

[0029] Constraints: ω1+ω2+ω3=1, ω1>0, ω2>0, ω3>0;

[0030] Where T is the number of days to be predicted, N is the number of combined prediction models (3), and ω i Let L be the weight of the i-th model. it M it R it Let be the actual values ​​of the minimum, average, and maximum electricity load on day t for the i-th model, respectively. These are the predicted minimum, average, and maximum power load values ​​for the i-th model on day t, respectively.

[0031] 43) By performing a perfect square operation on the objective function E, the problem is transformed into a quadratic linear programming problem, which is then converted into solving a problem that satisfies the condition minE. 2 The weight matrix W is obtained, which gives the triangular fuzzy optimal combination prediction model, as follows:

[0032]

[0033]

[0034] in,

[0035] M is the coefficient matrix of the standard quadratic programming form constructed from actual data and predicted data.

[0036] in,

[0037] The process of obtaining the triangular fuzzy prediction results for the power load includes the following steps:

[0038] 51) Taking the predicted power load data of 11 sets of XGBoost model, MSVR model, and MLP model and their corresponding actual data as inputs, respectively, the software is used for optimization and solution to obtain an 11×3 weight matrix W, W=(ω ij ) 11×3 Where, ω ij Let be the weights of the j-th model under the i-th group of intrinsic mode functions;

[0039] 52) Assign weights w to the individual model prediction results of each group of intrinsic modulus functions. ij Linear combination is performed to obtain the triangular fuzzy optimal combination prediction result of each set of intrinsic mode functions;

[0040] 53) The 11 sets of prediction results are linearly superimposed to obtain the integrated triangular fuzzy prediction result of power load, that is, the power load prediction result is obtained.

[0041] Beneficial effects

[0042] The present invention provides a triangular fuzzy prediction method for power load based on MEEMD and optimal combination integration. Compared with the prior art, this method performs empirical mode decomposition on the raw power load data, constructs triangular fuzzy numbers, and uses three models—XGBoost, MSVR, and MLP—for combined prediction. It improves the prediction accuracy and disperses the prediction risk by utilizing the MEEMD algorithm and optimal combination integration technology.

[0043] In the field of power forecasting, this invention improves the forecasting performance by 72.4%, 64.0%, and 63.0% respectively compared to using XGBoost, MSVR, and MLP models alone. This greatly improves the accuracy of forecasting and can better reflect the multi-dimensional information of power load data, thereby strengthening the country's management and control of power load. Attached Figure Description

[0044] Figure 1 This is a sequence diagram of the method of the present invention;

[0045] Figure 2 An intrinsic model decomposition diagram for predicting the maximum daily power load for the last 30 days;

[0046] Figure 3 An intrinsic model decomposition diagram for predicting the daily minimum electricity load for the last 30 days;

[0047] Figure 4 An intrinsic model decomposition diagram for predicting the daily average electricity load for the last 30 days;

[0048] Figure 5 To predict the daily maximum power load for the last 30 days using various models and combined models, and the original sequence of line graphs;

[0049] Figure 6 The model and combined model are used to predict the minimum daily power load for the last 30 days, and the original sequence is plotted as a line graph.

[0050] Figure 7 To predict the daily average electricity load of each model and combined model for the last 30 days, and to plot the original sequence.

[0051] Figure 8 To predict the undecomposed models for the last 30 days, and to create a line plot of the daily maximum electricity load from the original sequence;

[0052] Figure 9 A line plot of the undecomposed models and the original sequence of daily minimum electricity loads for the last 30 days;

[0053] Figure 10 To predict the last 30 days, we use individual models without decomposition and a line plot of the original sequence's daily average electricity load. Detailed Implementation

[0054] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0055] like Figure 1 As shown, the present invention provides a triangular fuzzy prediction method for power load based on MEEMD and optimal combination integration, comprising the following steps:

[0056] The first step is to obtain the dataset: obtain hourly power load data on a daily basis, and calculate the maximum value R, minimum value L, and average value M of the daily power load.

[0057] Here, using actual data as an example, the electricity load data comes from the website www.wia.gov / , specifically hourly electricity load data for California, USA, from July 1, 2015 to August 26, 2022. This data is obtained by the US hourly grid monitor and has strong timeliness and reliability. California, being the most populous state in the US, has a huge demand for electricity, and the contradiction between power supply and demand is becoming increasingly prominent, with the peak load of the grid continuously increasing. Therefore, the monitoring and regulation of its electricity load has profound research significance. Analysis of the obtained data shows that electricity load varies greatly within a 24-hour period. Generally speaking, electricity consumption at night, especially late at night, is much lower than during the day. The time with the highest demand for electricity is called "peak" time, and the rest is "off-peak" time. The marginal cost of supplying electricity during peak hours is higher because all equipment operates at full capacity; while the marginal cost during off-peak hours is lower because only the most efficient generators are operating. Therefore, simultaneously predicting the maximum value R, minimum value L, and average value M of the electricity load has profound research significance. The following formula is used to calculate the daily electricity load R, M, and L:

[0058]

[0059] R = max(e h ),h=1,2,3K 24;

[0060] L = min(e h ),h=1,2,3K 24;

[0061] Where e h The electrical load at time h.

[0062] The second step is the decomposition of the dataset: the power load data is decomposed using the MEEMD method to obtain 11 sets of intrinsic mode functions that eliminate the mixing phenomenon.

[0063] Because power load data is nonlinear, random, and time-varying, a triangular fuzzy prediction method for power load based on MEEMD and optimal combination integration is needed. This method involves empirical mode decomposition, fuzzification, and integration of the original power load information, and then applying it to the field of power forecasting. Multivariate Empirical Mode Decomposition (MEEMD) is an improved adaptive data analysis method that can obtain effective information from power load data. It reduces the impact of mode aliasing by adding auxiliary white noise and decomposes the time series into several intrinsic mode functions and a residual, forming an EEMD decomposition. Simultaneously, based on EEMD, the R, M, and L parameters of the power load data are decomposed to form the MEEMD decomposition.

[0064] To eliminate the mixing phenomenon of MEMD, this method adds auxiliary white noise to the original data for decomposition. Since the original electricity consumption data is of a large order of magnitude, it is first necessary to divide the original data by its own standard deviation to eliminate the influence of order of magnitude and dimensions. Then, white noise with a mean of 0 and a standard deviation of NTSD = 0.05 is added to the original data to obtain a new set of data containing white noise. Next, empirical mode decomposition is performed on the data containing white noise to obtain 11 sets of intrinsic mode functions (IMF1). 1～11 The extra computation number Ne = 50 was set, and different randomly generated white noise was added in each decomposition. Finally, a total of 11 sets of intrinsic mode functions (IMF1 to IMF2) containing white noise were obtained. 51 ) 1～11 Then, sum the intrinsic mode functions containing white noise in each set and divide by Ne+1 to obtain the mean function of the intrinsic mode functions containing white noise in that set. Then multiply by the standard deviation to restore the order of magnitude, and finally obtain the 11 sets of intrinsic mode functions IMF1 to IMF11 that eliminate the mixing phenomenon.

[0065] The formula for adding white noise is: (R,M,L)÷std(R,M,L)+randn*NTSD, where randn is a random number following a standard normal distribution; the formula for calculating the final intrinsic mode function is: Next, the R, L, and M of each set of intrinsic mode functions are combined to form a triangular fuzzy number A(R, L, M). Here, a three-dimensional variable consisting of the maximum, minimum, and mean values ​​is used for prediction. This is to increase the accuracy of the model prediction and to make the prediction results more comprehensive by making more comprehensive use of the existing data information. Next, the dataset needs to be constructed. The data of m days is used to form an n-day dataset (n = m-1). The first k days are selected as the training set and the last nk days are selected as the test set. The specific construction method is to divide the m-day data (2613 days) into a dataset of size n (2612 days) by using the method of predicting the next day from the previous day. Next, in order to facilitate the training and testing of the model, the divided dataset is divided into a k-day training set and an nk-day test set, where nk = 30, in order to prepare for the prediction of the power load in the last 30 days.

[0066] The decomposition of the dataset includes the following steps:

[0067] (1) Divide the original data by its own standard deviation to eliminate the influence of order of magnitude and dimension, thus preparing for the subsequent superposition of white noise.

[0068] (2) Add a normal white noise with a standard deviation of NTSD and a mean of 0 to the original data to generate a new set of data containing white noise.

[0069] (3) Then, empirical mode decomposition was performed on the data containing white noise to obtain 11 sets of intrinsic mode functions.

[0070] (4) Set the additional number of calculations to Ne, and repeat steps (3) and (4) for Ne times to obtain a total of 11 sets of intrinsic mode functions containing white noise.

[0071] (5) Sum the intrinsic mode functions containing white noise in each set and divide by Ne+1 to obtain the mean of the intrinsic mode functions containing white noise in that set. Then multiply by the standard deviation to restore the order of magnitude, and then obtain 11 sets of intrinsic mode functions to eliminate the mixing phenomenon.

[0072] The third step involves single-model prediction: the first k data points are used as the training set, and the last nk data points are used as the test set. XGBoost, MSVR, and MLP are used to predict the electricity load data for the last nk days of each intrinsic mode function. First, the three machine learning models are used to perform single-model prediction on the dataset after MEEMD decomposition. Then, the same three machine learning models are used to perform single-model prediction on the dataset without MEEMD decomposition to form a control group. The three machine learning models are XGBoost (eXtreme Gradient Boosting), MSVR (Multioutput Support Vector Rregression), and MLP (Multilayer Perceptron) to predict the electricity load data.

[0073] a. Extreme gradient boosting model

[0074] XGBoost is a tree-based method that combines ensemble learning with ensemble learning. Its basic tree structure is the Classification and Regression Tree (CART). Similar to locally weighted linear regression, tree-based regression algorithms are also a type of local regression algorithm, which models the data separately by dividing the dataset into multiple parts. In this invention, the model sets the number of weak evaluators in the ensemble to 300 to improve the model's learning ability. Setting it too high may also affect the model's performance; therefore, 300 is optimal. The maximum depth of each tree is set to 5, which allows for a good fit to the training set data and prevents overfitting during model training. The sampling ratio for each evaluator is set to 0.8, meaning that 80% of the samples are selected from the 2582 training samples to build the tree model each time. The feature sampling ratio during tree construction is also set to 0.8, which effectively prevents overfitting of each evaluator. The objective function used in this model is a linear regression loss function, which is used to minimize the mean squared error for regression modeling. The objective function of this model is as follows: Where Ω(f) t ) is a penalty item. This structure has two parts: γT, which controls the tree structure, and the regularization term. The number of leaves, T, represents the entire tree structure because all trees in XGBoost are CART trees (binary trees), so we can determine the tree depth based on the number of leaves. γ is a custom parameter that controls the number of leaves. As for the second part, the regularization term, L2 regularization is used here, where λ controls the regularization strength. When λ is 0, the objective function is the same as that of a normal gradient boosting tree, f. t (x i ) represents one of the regression trees, and l represents the loss function, such as the squared loss function l(y i ,y i )=(y i -y i ) 2 .

[0075] b. Multivariate Support Vector Machine Model

[0076] Support Vector Machines (SVMs) can effectively handle regression and pattern recognition problems, thus having application value in the field of monitoring and prediction. The SVM regression classification idea is to map linearly inseparable data to a higher-dimensional space using a kernel function. This invention utilizes a multivariate SVM model to predict power load data, where the feature dimensions are three-dimensional: maximum power load R, minimum power load L, and average power load M. A Gaussian kernel function is used, which can effectively predict this dataset. The decision function of this model is shown in the following equation:

[0077]

[0078] Where, α i Let x be a Lagrange multiplier, b be the deviation value, and x be a Lagrange multiplier. i y i These are support vectors in binary classification, K(x·x) i ) is the kernel function.

[0079] c. Multilayer Perceptron Model

[0080] A Multilayer Perceptron (MLP) is a feedforward artificial neural network that maps a set of input vectors to a set of output vectors, exhibiting significant learning and reasoning capabilities. It is well-suited for triangular fuzzy prediction of power load. Its input layer has 3 neurons, representing R (maximum power load), M (average power load), and L (minimum power load), and several hidden layers. This invention uses 130 hidden layer neurons, and by altering the linearity rule through an activation function, it overcomes the perceptron's weakness in recognizing linearly inseparable data, allowing the neural network to be applied to more nonlinear models. The output layer has 3 neurons, representing the predicted three-dimensional R, M, and L results for the next day.

[0081] The single-model prediction includes the following steps:

[0082] (1) Set up the XGBoost model using the sklearn library function in Python. The parameters are set as follows: the ensemble learning rate is set to 0.1, the number of weak evaluators in the ensemble is set to 300, the maximum depth of each tree is set to 5, the sum of the second derivatives on the leaf nodes is similar to the sample weights being set to 1, the proportion of samples drawn during random sampling is set to 0.8, the proportion of features randomly sampled each time a tree is generated is set to 0.8, and the L2 regularization term is used, with the regularization term parameter lambda set to 1.

[0083] (2) Set up the MSVR model, using the Gaussian kernel function rbf. The parameters sigma that affect the radial range of rbf are set to 560000000, the penalty coefficient C is set to 580000, the error precision for stopping training is set to 1e-10, and Epsilon is set to 1.04. Specify epsilon-tube, which contains points in the training loss function that are not penalized and whose distance from the actual value is less than the Epsilon prediction.

[0084] (3) Set up the MLP model, using the ReLU activation function, with 3 neurons in the input layer, 130 neurons in the hidden layer, and 3 neurons in the output layer.

[0085] (4) Train the XGBoost model, MSVR model and MLP model respectively with the power load dataset of the first k days, and then train the power load dataset of the next nk days with the above three models respectively to make single-item model predictions and obtain the prediction results of the above three models.

[0086] The fourth step is to construct a triangular fuzzy optimal combination prediction model: using the prediction results of the three individual prediction models, construct a triangular fuzzy optimal combination prediction model for each set of intrinsic mode functions that minimizes the sum of squared errors in power load prediction.

[0087] Different intrinsic mode functions (IMF) components have different data distribution characteristics and variation patterns. The three models mentioned in the above steps have varying prediction effects on the 11 sets of IMFs obtained. Some IMF components are suitable for the MSVR prediction model, while others are suitable for the XGBoost and MLP models. Therefore, this invention applies different IMF components to a combined prediction model with different weights, which can significantly improve the prediction effect of power load compared to single-item prediction models. Next, this invention will focus on the triangular fuzzy optimal combined prediction model.

[0088] First, we define the prediction errors of the three endpoints L, M, and R in the triangular fuzzy number, which reflect the prediction effects of the minimum, mean, and maximum values. Here, we consider the errors of the three endpoints to be equally important, and thus obtain the prediction error of the triangular fuzzy number.

[0089] Then, the objective function for triangular fuzzy optimal combination optimization and its constraints are defined as follows:

[0090]

[0091] Constraints: ω1+ω2+ω3=1, ω1>0, ω2>0, ω3>0;

[0092] Where T is the number of days to be predicted (2612-k), N is the number of combined prediction models (3), and ω i Let L be the weight of the i-th model. it M it R it Let be the actual values ​​of the minimum, average, and maximum electricity load on day t for the i-th model, respectively. These are the minimum, average, and maximum predicted values ​​of the power load on day t for the i-th model, respectively.

[0093] Finally, to facilitate the solution, the objective function is transformed into a quadratic linear programming problem by performing a perfect square operation on E. Therefore, the solution is transformed into finding the solution that satisfies minE. 2 The weight matrix W is used to obtain the triangular fuzzy optimal combination prediction model, as follows:

[0094]

[0095]

[0096] in,

[0097] M is the coefficient matrix of the standard quadratic programming form constructed from actual data and predicted data.

[0098] in,

[0099] The fifth step is to obtain the triangular fuzzy prediction results of the power load: the software is used to solve the optimal combination prediction model of the triangular fuzzy function for each set of intrinsic mode functions, and the prediction results of 11 sets of intrinsic mode functions are obtained. Then, the prediction results are linearly superimposed to obtain the integrated prediction results, that is, the power load prediction results.

[0100] First, the predicted power load data from 11 sets of XGBoost, MSVR, and MLP models, along with their corresponding actual data, are used as inputs (corresponding to 11 IMF components). Then, the MATLAB optimization toolbox is used for optimization to obtain the optimal weight matrix W of order 11×3, where W = (ω ij ) 11×3 Where, ω ij Let be the weights of the j-th model under the i-th group of intrinsic mode functions;

[0101] Then, the individual model prediction results of each group of intrinsic modulus functions are assigned weights ω. ij A linear combination is performed to predict the power load data for the last nk days (n=2612). The triangular fuzzy optimal combination prediction results for each set of intrinsic mode functions are obtained;

[0102] Finally, the 11 sets of prediction results are linearly superimposed to obtain the integrated combined prediction result, thus obtaining the triangular fuzzy prediction result of power load.

[0103] like Figure 2 , Figure 3 , Figure 4 The figures shown are the intrinsic modulus function decomposition diagrams of the predicted daily maximum, minimum, and average electricity load for the last 30 days. Figure 2 , Figure 3 , Figure 4 The data shows the daily maximum, minimum, and average power load variation patterns over time for the original sequence and 11 sets of intrinsic mode functions. The intrinsic mode functions that eliminate the mixing phenomenon are obtained through MEEMD decomposition, thus making sufficient preparations for subsequent predictions.

[0104] like Figure 5 , Figure 6 , Figure 7 The figures shown are line graphs of the daily maximum, minimum, and average electricity load for the prediction of the last 30 days, for each model and combined model, and for the original sequence. Figure 5 , Figure 6 , Figure 7 As can be seen, the combined power load prediction model has the highest fit with the original sequence, and both are stronger than the prediction performance of the single-item model.

[0105] like Figure 8 , Figure 9 , Figure 10The figures shown are line graphs of the daily maximum, minimum, and average electricity load for the last 30 days, presented by the individual undecomposed models and combined models, compared with the original sequence. Figure 5 and Figure 8 , Figure 6 and Figure 9 , Figure 7 and Figure 10 As can be seen, the maximum, minimum, and average power load data after MEEMD decomposition are better than the prediction methods without decomposition, whether used for single-item model prediction or combined model prediction.

[0106] In summary, this invention constructs a triangular fuzzy number A(R, M, L) from power load data, which effectively reflects the multi-dimensional information of the power load data and improves prediction performance. The MEEMD method used in this invention decomposes complex and irregular raw power load data, obtaining 11 sets of intrinsic modulus functions (IMFs) containing local features of different time scales of the original sequence data. By predicting each IMF set separately and then combining the prediction results, it is possible to effectively and accurately predict each component feature, thereby improving the final prediction performance. MSVR, MLP, and XGBoost models have varying degrees of effectiveness in predicting power load data for different IMF components. Combining these models effectively combines their advantages, enabling good predictions for different sequence data and thus improving prediction performance. This invention integrates the MEEMD method with optimal combination for triangular fuzzy prediction, greatly improving prediction accuracy and allowing for multi-dimensional analysis of the prediction results, thereby strengthening national management and control of power load.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A power load triangular fuzzy prediction method based on MEEMD and optimal combination integration, characterized in that, The method comprises the following steps: 11) data set acquisition: acquire hourly power load data in units of days, and calculate the maximum value R, minimum value L and average value M of the power load of each day respectively; 12) data set decomposition: decompose the power load data using the MEEMD method to obtain 11 groups, i.e. 11*3 intrinsic mode functions eliminating frequency mixing phenomenon; 13) single model prediction: the first k data are used as a training set, and the last n-k data are used as a test set, and XGBoost, MSVR and MLP are used to predict the power load data of the last n-k days of each group of intrinsic mode functions; 14) construction of a triangular fuzzy optimal combination prediction model: the prediction results of the three single prediction models are used to construct a triangular fuzzy optimal combination prediction model of each group of intrinsic mode functions, which minimizes the sum of squares of power load prediction errors; The triangular fuzzy optimal combination prediction model comprises the following steps: 141) define the prediction errors of the L, M and R endpoints in the triangular fuzzy number, which reflect the prediction effect of the minimum value, average value and maximum value, and assume that the errors of the three endpoints are equally important to obtain the prediction error of the triangular fuzzy number; 142) define the triangular fuzzy optimal combination optimization objective function and its constraint condition as follows: Constraint condition, ω1+ω2+ω3=1, ω1>0, ω2>0, ω3>0; where T is the predicted days, N is the number of combined prediction models 3, ω i is the weight of the i-th model, L it , M it , R it are the minimum, average and maximum actual values of the power load of the i-th model on the t-th day, respectively, are the minimum, average and maximum predicted values of the power load of the i-th model on the t-th day, respectively. 143) The optimization objective function E is converted into a quadratic linear programming by making the complete square processing, so it is converted into solving minE 2 The weight matrix W is obtained, that is, the triangular fuzzy optimal combination prediction model is obtained, as follows: wherein M is a coefficient matrix of a standard form of quadratic programming constructed by actual data and predicted data, wherein, 15) obtaining of the triangular fuzzy prediction result of the power load: the triangular fuzzy optimal combination prediction model of each group of intrinsic mode functions is solved by using software to obtain the prediction results of the 11 groups of intrinsic mode functions, and then linear superposition is performed to obtain the integrated prediction result, i.e. the power load prediction result.

2. The MEEMD and optimal combination integrated power load triangular fuzzy prediction method according to claim 1, characterized in that, The data set decomposition comprises the following steps: 21) divide the original data by the standard deviation thereof to eliminate the influence of the order of magnitude and dimension, and further prepare for subsequent superposition of white noise; 22) append normal distribution white noise with a standard deviation of NTSD and a mean value of 0 to the original data to generate a new group of data containing white noise; 23) further perform empirical mode decomposition on the data containing white noise to obtain 11 groups of intrinsic mode functions; 24) set the additional calculation number as Ne, repeat steps 23) and 24) for Ne times, and finally obtain 11 groups of intrinsic mode function sets containing white noise; 25) sum each group of intrinsic mode function sets containing white noise and divide by Ne+1 to obtain the mean function of the intrinsic mode functions containing white noise in the set, and then multiply by the standard deviation to restore the order of magnitude, and further obtain 11 groups of intrinsic mode functions eliminating frequency mixing phenomenon.

3. The MEEMD and optimal combination integrated power load triangular fuzzy prediction method according to claim 1, characterized in that, The single model prediction comprises the following steps: 31) Set the XGBoost model, use the sklearn library function of python, the set parameters are as follows: the integrated learning rate is set to 0.1, the number of weak evaluators in the integrated learning is set to 300, the maximum depth of each tree is set to 5, the sum of the second derivative on the leaf node, that is, the sample weight is set to 1, the sample ratio when random sampling is set to 0.8, the ratio of randomly sampled features when generating a tree is set to 0.8, L2 regularization term is used, and the parameter lambda of the regularization term is set to 1; 32) Set the MSVR model, use the Gaussian kernel function rbf, set the parameter sigma that affects the radial action range of rbf to 560000000, set the penalty coefficient C to 580000, set the error precision of stopping training to 1e-10, set Epsilon to 1.04, specify epsilon-tube, and epsilon-tube contains points predicted by the distance between the actual value and the distance less than Epsilon in the training loss function without penalty; 33) Set the MLP model, use the sklearn library function of python, the set parameters are as follows: use the relu activation function, the number of input layer neurons is 3, the number of hidden layer neurons is set to 130, and the number of output layer neurons is 3; 34) The power load data set of the previous k days is respectively brought into the XGBoost model, the MSVR model and the MLP model for training, and then the power load data set of the next n-k days is respectively brought into the above three kinds of models for single model prediction, and the prediction results of the above three kinds of models are obtained.

4. The MEEMD and optimal combination integrated power load triangular fuzzy prediction method according to claim 1, characterized in that, The obtaining of the power load triangular fuzzy prediction result comprises the following steps: 41) Take the prediction results of the 11 groups of power load data of XGBoost model, MSVR model and MLP model respectively and the corresponding actual data as input, and use software to optimize and solve to obtain the 11x3 order weight matrix W, W=(ω ij ) 11×3 where ω ij is the weight of the jth model under the ith intrinsic mode function. 42) assigning a weight w to the single-model prediction result of each group of eigenmode functions ij performing linear combination, and thus obtaining the triangular fuzzy optimal combination prediction result of each group of eigenmode functions; 43) Linearly superimpose 11 groups of prediction results to obtain an integrated power load triangular fuzzy prediction result, that is, obtain the power load prediction result.

Images