Regional power load prediction method of EEMD-LASSO-GM model
Through the EEMD-LASSO-GM model, the problems of traditional power load prediction methods missing long-sequence memory and insufficient potential feature mining when processing nonlinear, multi-factor complex data are solved, and more accurate and reliable power load prediction is achieved.
Patent Information
- Application Number
- CN202510031482.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-09
- Publication Date
- 2025-05-30
AI Technical Summary
When traditional power load prediction methods deal with nonlinear, multi-factor complex power load data, they are prone to problems such as long-sequence memory loss, messy data structure and insufficient potential feature mining, which affects the accuracy of the prediction results.
The EEMD-LASSO-GM model is used to perform EMD decomposition through the EEMD model to deeply explore the potential characteristics in the power load data; the LASSO model is used to screen the factor database data to reduce data redundancy and correlation; finally, the GM model uses the screened key factors to optimize the model to achieve accurate prediction of future power loads.
It improves the accuracy and reliability of power load prediction, can more effectively explore potential characteristics in the data, reduce the computing power burden during the prediction process, and meet the prediction accuracy of different needs.
Smart Images
Figure CN120073665A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to power system planning, and particularly to a regional power load forecasting method based on the EEMD-LASSO-GM model. Background Art
[0002] In a power system, load forecasting is a crucial task that is directly related to multiple aspects such as power grid planning, dispatching optimization, and energy management. Among them, common load forecasting methods include various regression methods and time series analysis methods, etc. When facing complex power load data with high nonlinearity and intertwined multi-factors, these methods often face the following problems:
[0003] First, power load data usually has a long time span and is affected by multiple factors such as seasonal changes, weather conditions, holiday effects, and economic development status, showing a high degree of dynamics and uncertainty. When dealing with such long-sequence data, traditional methods may have problems of sequence memory loss due to huge computational amounts and increased model complexity, thus affecting the accuracy of future load trend forecasting.
[0004] Second, the internal of power load time series data often contains rich potential features and patterns, such as inherent periodicity and trend. Due to the limitations of traditional methods in dealing with complex nonlinear relationships, it is difficult to deeply mine these potential features, resulting in a large deviation between the prediction result and the actual load, thus limiting the accuracy of future regional load forecasting.
[0005] Therefore, for the prediction of nonlinear and multi-factor complex power loads, traditional load forecasting methods face problems such as long-sequence memory loss, data structure disorder, and insufficient mining of potential features in practical applications. The existence of these problems affects the accuracy of the prediction results when performing regional power load forecasting.
[0006] Thus, how to improve the accuracy and reliability of load forecasting has become an inevitable problem in the current research in the power field. Summary of the Invention
[0007] To solve the problems of long-sequence memory loss, data structure disorder, and insufficient mining of potential features faced by traditional load forecasting methods in the prediction of nonlinear and multi-factor complex power loads as mentioned in the above background art, the present invention provides the following technical solutions:
[0008] A regional electric load forecasting method based on the EEMD-LASSO-GM model, including the EEMD-LASSO-GM model for forecasting the electric load in the area to be predicted and the database required by this model. The EEMD-LASSO-GM model is jointly constructed by the EEMD model, the LASSO model and the GM model, and the database includes a historical database, a noise database, a factor database and a key factor database.
[0009] Regarding the database:
[0010] The historical database is the time series data composed of the actual values of the electric load in the area to be predicted in the past. The noise database is a set of white noises preset to change the extreme point characteristics of the historical database. The factor database is the intrinsic mode component data derived from the EMD decomposition of the superposition of the data in the historical database and the data in the noise database by the EEDM model. The key factor database is the data retained by screening the data in the factor database by the LASSO model.
[0011] Regarding the model operation:
[0012] The EEMD model obtains the data in the historical database and the data in the noise database, superimposes them to generate the signal to be decomposed, and obtains several intrinsic mode components through EMD decomposition to derive the factor database. In addition, the EEMD model obtains the data in the key factor database and performs an average operation on this data. The LASSO model obtains the data in the factor database, extracts the key factors affecting the electric load in the area to be predicted to derive the key factor database. The GM model is constructed using the key factor database and optimized using the least squares method to judge the future changes in the electric load in the area to be predicted.
[0013] Further, the complete data of the time series data in the historical database is X(T), and the sequence segment used for data analysis by the EEMD model is x(t), and the data structure is as follows:
[0014] X(T) = {x(1), x(2), …, x(T)}
[0015] x(t) = {x(1 + n), x(2 + n), …, x(T - m)}, x(t) ∈ X(T)
[0016] In the formula, x(T) represents the actual value of the electric load in the area to be predicted at time T; both n and m are natural numbers.
[0017] Further, all the white noise signal sets in the noise database are WN(T), and the white noise signal set used for data analysis by the EEMD model is wn(t), and the data structure is as follows:
[0018] WN(T) = {wn1 (t), wn 2 (t), …, wn i (t)}
[0019] wn(t) = {wn 1+n (t), wn 2+n (t), …, wn i-m (t)}, wn(t) ∈ WN(T)
[0020] Wherein, wn i (t) represents the i-th white noise signal, where i is a non-zero natural number; both n and m are natural numbers.
[0021] Furthermore, the operation steps of the EEMD model are as follows:
[0022] S1.1. Generation of the signal to be decomposed. That is, according to the sequence segment data x(t) selected by the decision maker, perform the action of selecting a single white noise signal wn i (t) from the white noise signal set wn(t) and superimposing it on the sequence segment data x(t) to generate the signal to be decomposed x i (t), and the white noise signal wn i (t) selected in each superimposing action is different.
[0023] S1.2. Derivation of the factor database. That is, perform EMD decomposition on each signal to be decomposed x i (t) generated in the previous step to obtain their respective intrinsic mode component data to jointly form the data set C i (t) of the factor database.
[0024] S1.3. Derivation of the key factor database. That is, use the LASSO model to obtain key factors from the factor database to form the key factor database.
[0025] S1.4. Calculation of the key factor average value. That is, obtain the data of the key factor database and perform an average operation to obtain the key sequence g 0 (t) for GM model establishment.
[0026] Furthermore, the data set of the factor database is C i (t), and the data structure is as follows:
[0027] Ci(t) = {c 1 (t), c 2 (t), …, c i (t)}
[0028] c i (t) = {c i1 (t), c i2(t), …, c ij (t)}
[0029] wherein, c i (t) represents the set of intrinsic mode components obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; c ij (t) represents the j-th intrinsic mode component obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; both i and j are non-zero natural numbers.
[0030] Furthermore, the operation steps of the LASSO model are as follows:
[0031] S2.1. Preset the threshold value Z. Obtain its absolute value |Z|.
[0032] S2.2. Obtain the extreme value information of each intrinsic mode component c i (t) within each set of intrinsic mode components c ij (t). That is, obtain the maximum value of the intrinsic mode component c ij (t) to obtain its average value Z iMAX , and obtain the minimum value of the intrinsic mode component c ij (t) to obtain its average value Z iMIN .
[0033] S2.3. Factor screening and derivation of key factors. That is, separately compare the extreme value information of each intrinsic mode component c ij (t) with the preset threshold value, and screen out all the intrinsic mode components that meet the screening conditions within each set of intrinsic mode components c i (t). Then, the factors retained within each set of intrinsic mode components c i (t) are key factors, and the sets of intrinsic mode components that only retain key factors together constitute the key factor database G i (t).
[0034] Furthermore, the data set of the key factor database is as follows, and the data structure is as follows:
[0035] G i (t) = {g 1 (t), g 2 (t), …, g i (t)}, G i (t) ∈ C i (t)
[0036] g i (t) = {g i1 (t), g i2 (t), …, g ij (t)}, g i (t) ∈ ci (t)
[0037] In the formula, g i (t) represents the set of intrinsic mode components that only retain key factors obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; g ij (t) represents the j-th intrinsic mode component in the set of key factors obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; both i and j are non-zero natural numbers.
[0038] Furthermore, the establishment and operation steps of the GM model are as follows:
[0039] S3.1. Input and derivation of the original sequence, that is, taking the key sequence g 0 (t) as the original sequence for a single accumulation to derive the generated sequence g 1 (t).
[0040] S3.2. Construct a differential equation, that is, establish a differential equation based on the data of the generated sequence g 1 (t).
[0041] S3.3. Optimization of the differential equation, that is, using the least squares method to optimize the differential equation constructed in the previous step.
[0042] S3.4. Make a prediction, that is, according to the optimized differential equation, predict the change of the power load in the area to be measured under the established conditions.
[0043] Compared with the prior art, the beneficial effects of the present invention are:
[0044] First, the EEMD-LASSO-GM model of the present invention first analyzes the time series data composed of the actual values of the power load in the area to be predicted in the past through the EMD decomposition of the EEMD model, deeply excavates the potential features hidden in the data, and provides reliable information support for subsequent predictions; it enriches the diversity of independent variables of the data before analysis and provides more flexibility for subsequent predictions.
[0045] Second, the EEMD-LASSO-GM model of the present invention can use the LASSO model to screen the hidden data mined by the EEMD model, thereby reducing the redundancy and correlation of the data, and further reducing the computing power burden in the prediction process and improving the accuracy of the prediction results. Among them, the screening conditions in the LASSO model adopt a changeable method, making the data screening more flexible, so as to meet different requirements for the accuracy of the prediction results.
[0046] III. For the data retained after the operation of the LASSO model in the EEMD-LASSO-GM model of the present invention, the EEMD model is used to perform ensemble averaging operation on it to condense the data sequence, providing conditions for constructing a real-time GM model in the later stage, and changing the subsequent constructed real-time GM model by changing this condition.
[0047] IV. In the EEMD-LASSO-GM model of the present invention, the GM model is constructed in real time based on the data obtained from the ensemble averaging operation of the EEMD model, and the least squares method is used for model optimization, so as to use the optimized GM model to predict the future power load in the area to be predicted. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 is the architecture diagram of the EEMD-LASSO-GM model of the present invention;
[0049] Figure 2 is the logic diagram of the EEMD model in the EEMD-LASSO-GM model of the present invention;
[0050] Figure 3 is the EMD decomposition logic diagram in the EEMD-LASSO-GM model of the present invention;
[0051] Figure 4 is the logic diagram of the LASSO model in the EEMD-LASSO-GM model of the present invention;
[0052] Figure 5 is the logic diagram of the GM model in the EEMD-LASSO-GM model of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0053] The preferred specific embodiments for implementing the present invention are described in detail below, and a clear and complete description is made in conjunction with the accompanying drawings.
[0054] Please refer to Figures 1 - 5 , the present invention provides a method for predicting regional power load of the EEMD-LASSO-GM model. This method includes a model side and a database side. Through the information interaction between the model side and the database side, the model side is used to predict the future change of power load in the area to be predicted.
[0055] Specifically, on the model side, that is, the EEMD-LASSO-GM model, an EEMD model, a LASSO model, and a GM model are built therein; on the database side, that is, the database used by the EEMD-LASSO-GM model, it contains raw data and derived data, which are specifically divided into a historical database, a noise database, a factor database, and a key factor database. Among them, the raw data refers to the data that has been entered into the database without being operated by the EEMD-LASSO-GM model, and the derived data refers to the data obtained after the EEMD-LASSO-GM model operates on the raw data.
[0056] In the present invention:
[0057] The data in the historical database is raw data, which is a time series data composed of the actual values of the power load in the past time of the area to be predicted. Its data structure is as follows:
[0058] X(T) = {x(1), x(2), …, x(T)}
[0059] In the formula, X(T) represents all the data entered into the historical database; x(T) represents the actual value of the power load at time T in the past time of the area to be predicted.
[0060] In the application of the present invention:
[0061] Before performing the EEMD model operation, the decision maker who conducts the prediction behavior selects the data as samples from the historical database according to the requirements. The data structure of the sample is as follows:
[0062] x(t) = {x(1 + n), x(2 + n), …, x(T - m)}, x(t) ∈ X(T)
[0063] In the formula, x(t) represents the data selected by the decision maker from the historical database as samples, and both n and m are natural numbers.
[0064] Preferably, for predicting the change of the power load in the current month or the next few months of the area to be predicted, the decision maker can start selecting from the actual value of the power load in the previous month of the area, and successively select the time series data of the actual values of the power load for several months in the time backtracking direction. These selected data together constitute the sample data x(t) in the historical database. Among them, the specific time length of the sample data x(t) is determined by the decision maker according to the needs.
[0065] Preferably, for predicting the power load change in a certain future month of the area to be predicted, decision-makers can sequentially select the actual power load values of this month in previous years of this area from the historical database in the time-backtracking direction, and these selected data together constitute the sample data x(t) in the historical database. Among them, the specific time length of the sample data x(t) is determined by the decision-makers according to their needs.
[0066] Preferably, for predicting the power load change in this quarter or several future quarters of the area to be predicted, decision-makers can start selecting from the actual power load value of the previous quarter of this area, and sequentially select the time series data of the actual power load values of several quarters in the time-backtracking direction. These selected data together constitute the sample data x(t) in the historical database. Among them, the specific time length of the sample data x(t) is determined by the decision-makers according to their needs.
[0067] Preferably, for predicting the power load change in a certain future quarter of the area to be predicted, decision-makers can sequentially select the actual power load values of this quarter in previous years of this area from the historical database in the time-backtracking direction, and these selected data together constitute the sample data x(t) in the historical database. Among them, the specific time length of the sample data x(t) is determined by the decision-makers according to their needs.
[0068] Preferably, for predicting the power load change in this year or several future years of the area to be predicted, decision-makers can start selecting from the actual power load value of the previous year of this area, and sequentially select the time series data of the actual power load values of several years in the time-backtracking direction. These selected data together constitute the sample data x(t) in the historical database. Among them, the specific time length of the sample data x(t) is determined by the decision-makers according to their needs.
[0069] In the present invention:
[0070] The data in the noise database are raw data, which is a set composed of a plurality of preset white noise signals with normal distribution. Its data structure is as follows:
[0071] WN(T) = {wn 1 (t), wn 2 (t), …, wn i (t)}
[0072] In the formula, WN(T) represents all the data entered into the noise database; wn i (t) represents the i-th white noise signal, and i is a non-zero natural number.
[0073] In the application of the present invention:
[0074] Before performing the EEMD model operation, several data that are respectively superimposed with the sample data x(t) are selected from the noise database by the decision maker who conducts the prediction behavior, and their data structure is as follows:
[0075] wn(t) = {wn 1+n (t), wn 2+n (t), …, wn i-m (t)}, wn(t) ∈ WN(T)
[0076] In the formula, wn(t) represents the superimposed data selected by the decision maker from the noise database for superimposing the sample data x(t); both n and m are natural numbers.
[0077] Preferably, for the selection of each white noise signal in the superimposed data wn(t), the decision maker can choose the method of artificial active selection or random selection.
[0078] In the present invention:
[0079] The operation of the EEMD model is divided into two stages. The first stage is the EMD decomposition stage, and the second stage is the averaging operation stage.
[0080] In the first stage of the EEMD model:
[0081] The EEMD model obtains the data from the historical database and the data from the noise database, superimposes the two sets of data to generate the signal to be decomposed, and decomposes the signal to be decomposed through EMD to obtain several intrinsic mode components. The data set composed of these intrinsic mode components constitutes the data of the factor database.
[0082] Preferably, the specific steps of the first stage of the EEMD model are as follows:
[0083] S1.1. Generation of the signal to be decomposed. That is, the decision maker selects the sequence segment as the sample data x(t) from the complete data of the time series in the historical database, and selects the superimposed data wn(t) from the noise database. Subsequently, each white noise signal wn i (t) in the superimposed data wn(t) is respectively superimposed with the sample data x(t) to generate the signal to be decomposed x i (t), and the white noise signal wn i (t) selected in each superimposing action is different. Among them, the data structure of the signal to be decomposed x i (t) generated in this step is as follows:
[0084] x i (t) = {x 1 (t), x 2 (t), …, x i(t)}
[0085] Where x i (t) represents the signal to be decomposed generated by adding white noise to the sample data x(t) for the i-th time.
[0086] S1.2, Derivation of the factor database. That is, perform EMD decomposition on each signal x i (t) generated in step S1.1 to obtain the intrinsic mode component data corresponding to each signal x i (t), and these data together constitute the data of the factor database. Among them, the EMD decomposition process is as follows:
[0087] Process 1, Input the signal x i (t), and let r i (t) = x i (t). r i (t) is the residual.
[0088] Process 2, Find all the maximum points in the time series of r i (t) and use the cubic spline interpolation function to fit the upper envelope U i (t); find all the minimum points in the time series of r i (t) and use the cubic spline interpolation function to fit the lower envelope L i (t). Let UL i (t) = (U i (t) - L i (t)) / 2, and obtain h ij (t) = x i (t) - UL i (t).
[0089] Process 3, Judge whether h ij (t) meets the conditions of the intrinsic mode component. If it does not meet the conditions, let h ij (t) = r i (t), and enter Process 2; if it meets the conditions, enter Process 4. The conditions for meeting the intrinsic mode component are that the mean value of the upper and lower envelopes of h ij (t) is zero, and the sum of the number of its maximum points and the number of its minimum points is equal to or differs by one from the number of its zero-crossing points.
[0090] Process 4, Let h ij (t) = c ij (t), r i = x i (t) - c ij (t), j = j + 1. c ij(t) is the potential feature hidden in the data mined from the time series data composed of the actual power load values in the past time of the area to be predicted. These feature data together constitute the data of the factor database.
[0091] Process Five, judge r i (t) whether it meets the EMD stop condition. If it does not meet the condition, enter Process Two; if it meets, end this EMD decomposition. The stop condition is that r i (t) is a monotonic function.
[0092] Process Six, start a new round of EMD decomposition until all the signals to be decomposed generated are completely decomposed.
[0093] In the present invention:
[0094] The data of the factor database is derivative data, which is a data set jointly composed of each intrinsic mode component obtained by the EEMD model in the EMD decomposition process. Its data structure is as follows:
[0095] C i (t) = {c 1 (t), c 2 (t), …, c i (t)}
[0096] c i (t) = {c i1 (t), c i2 (t), …, c ij (t)}
[0097] In the formula, c i (t) represents the set of intrinsic mode components obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; c ij (t) represents the j-th intrinsic mode component obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; both i and j are non-zero natural numbers.
[0098] In the second stage of the EEMD model:
[0099] The LASSO model obtains the data of the factor database and screens this data. The data retained after screening constitutes the data of the key factor database and is used to substitute into the EEMD model for averaging operations.
[0100] Preferably, the specific steps of the second stage of the EEMD model are as follows:
[0101] S1.3. Derivation of the key factor database. That is, use the LASSO model to obtain key factors from the factor database to constitute the key factor database.
[0102] S1.4, Calculation of the average value of key factors. That is, obtain the data in the key factor database and perform an average operation to obtain the key sequence g 0 (t) for GM model establishment.
[0103] In the present invention:
[0104] The operation steps of the LASSO model are as follows:
[0105] S2.1, Preset threshold Z. That is, the decision maker obtains its absolute value |Z| by setting a value for comparing the extreme value information of the intrinsic mode component c ij (t).
[0106] S2.2, Obtain the extreme value information of each intrinsic mode component c i (t) in each set of intrinsic mode components c ij (t) in the factor database. That is, obtain the maximum value of the intrinsic mode component c ij (t) to obtain its average value Z iMAX , and obtain the minimum value of the intrinsic mode component c ij (t) to obtain its average value Z iMIN .
[0107] S2.3, Factor screening and derivation of key factors. That is, separately compare the extreme value information of each intrinsic mode component c ij (t) with the preset threshold, and screen out all the intrinsic mode components that meet the screening conditions in each set of intrinsic mode components c i (t). Then, the factors retained in each set of intrinsic mode components c i (t) are key factors, and the sets of intrinsic mode components that only retain key factors together constitute the key factor database G i (t).
[0108] In the present invention:
[0109] The data in the key factor database is derived data, which is the result of screening the data in the factor database by the LASSO model. The data structure is as follows:
[0110] G i (t) = {g 1 (t), g 2 (t), …, g i (t)}, G i (t) ∈ C i (t)
[0111] g i (t) = {g i1 (t), g i2 (t), …, gij (t), g i (t) ∈ c i (t)
[0112] Wherein, G i (t) represents all the data retained after screening all the data in the factor database by the LASSO model; g i (t) represents the set of intrinsic mode components retained after the sample data x(t) is added with white noise for the i-th time, decomposed by EMD, and screened by the LASSO model; g ij (t) represents the j-th intrinsic mode component in the set of key factors obtained by decomposing the sequence segment data x(t) by EMD after adding white noise for the i-th time; both i and j are non-zero natural numbers.
[0113] In the present invention:
[0114] The establishment and operation steps of the GM model are as follows:
[0115] S3.1. Input and derivation of the original sequence. That is, the key sequence g 0 (t) is used as the original sequence for a first-order accumulation to derive the generated sequence g 1 (t).
[0116] S3.2. Construct a differential equation. That is, a differential equation is constructed based on the data of the generated sequence g 1 (t).
[0117] S3.3. Optimization of the differential equation. That is, the differential equation constructed in the previous step is optimized using the least squares method.
[0118] S3.4. Make a prediction. That is, according to the optimized differential equation, the decision maker predicts the future power load change situation in the area to be measured based on the established conditions.
[0119] Preferably, the GM model is a GM(1,1) model.
[0120] Based on the above content and the accompanying drawings, those skilled in the art can understand and implement the present invention. In addition, any non-creative modifications made by those skilled in the art to the present invention without creative efforts still fall within the protection scope of the present invention.
Claims
1. A regional power load forecasting method based on EEMD-LASSO-GM model, characterized in that: A database including a historical database, a noise database, a factor database and a key factor database is established, and an EEMD-LASSO-GM model including an EEMD model, a LASSO model and a GM model is constructed, wherein: The historical database is the time series data consisting of the actual value of the power load in the past time in the area to be predicted; The noise database is a preset white noise set used to change the extreme point characteristics of the historical database; The EEMD model obtains data from the historical database and the noise database, superimposes them to generate the signal to be decomposed, and obtains several intrinsic modal components through EMD decomposition to derive the factor database. In addition, the EEMD model obtains data from the key factor database and performs an average operation on the data. The LASSO model obtains data from the factor database, extracts key factors that affect the power load in the area to be predicted, and derives a key factor database; The GM model is constructed using a database of key factors, and the least squares method is used to optimize the model to determine future power load changes in the area to be predicted.
2. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 1 is characterized in that: All the time series data in the historical database is X(T), and the sample data used for data analysis by the EEMD model is x(t). The data structure is as follows: X(T)={x(1),x(2),…,x(T)} x(t)={x(1+n), x(2+n),...,x(Tm)}, x(t)∈X(T) Where x(T) represents the actual value of the power load in the area to be predicted at time T; n and m are both natural numbers.
3. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 2 is characterized in that: The set of all white noise signals in the noise database is WN(T), and the set of white noise signals used for data analysis by the EEMD model is wn(t). The data structure is as follows: WN(T)={wn1(t)、wn2(t)、…、wn i (t)} wn(t)={wn 1+n (t)、wn 2+n (t)、…、wn i-m (t)},wn(t)∈WN(T) In the formula, wn i (t) represents the i-th white noise signal, i is a non-zero natural number; n and m are both natural numbers.
4. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 3 is characterized in that: The EEMD model operation steps are as follows: S1.
1. Generation of the signal to be decomposed, that is, selecting a single white noise signal wn from the white noise signal set wn(t) several times according to the sample data x(t) selected by the decision maker i (t) is superimposed on the sequence segment data x(t) to generate the signal to be decomposed x i (t) action, and each time the white noise signal wn selected in the superposition action i (t) different; S1.2, Derivation of factor database, that is, for each signal to be decomposed x generated in the previous step i (t) Perform EMD decomposition to obtain the intrinsic modal component data of each factor to form the data set C of the factor database. i (t); S1.3, derivation of key factor database, i.e., obtaining key factors from the factor database using LASSO model to form the key factor database; S1.
4. Calculate the average value of key factors, that is, obtain the data from the key factor database and perform average calculation to obtain the key sequence g 0 (t) is used for GM model establishment.
5. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 4 is characterized in that: The data set of the factor database is C i (t), the data structure is as follows: Ci(t)={c1(t)、c2(t)、…、c i (t)} c i (t)={c i1 (t)、c i2 (t)、…、c ij (t)} In the formula, c i (t) represents the set of intrinsic modal components obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; c ij (t) represents the jth intrinsic mode component obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; Both i and j are non-zero natural numbers.
6. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 5 is characterized in that: The LASSO model operation steps are as follows: S2.
1. Preset a threshold value Z and obtain its absolute value |Z|; S2.
2. Obtain the set c of each intrinsic modal component in the factor database i (t) Each natural modal component c ij (t), that is, to obtain the extreme value information of the intrinsic modal component c ij (t) to obtain its average value Z iMAX , get the intrinsic modal component c ij The minimum value of (t) is used to obtain its average value Z iMIN ; S2.3, Factor screening and derivation of key factors, that is, the natural modal components c ij The extreme value information of (t) is compared with the preset thresholds separately to filter out the set of intrinsic modal components c i (t) all the intrinsic modal components that meet the screening conditions, then the set of intrinsic modal components c i The factors retained in (t) are the key factors, so that the inherent modal component sets that only retain the key factors together constitute the key factor database G i (t).
7. The regional power load forecasting method according to the EEMD-LASSO-GM model of claim 4 or 6, characterized in that: The data set of the key factor database is, and the data structure is as follows: G i (t)={g1(t)、g2(t)、…、g i (t)},G i (t)∈C i (t) g i (t)={g i1 (t)、g i2 (t)、…、g ij (t)},g i (t)∈c i (t) In the formula, g i (t) represents the set of intrinsic modal components that retain only the key factors after white noise is added to the sequence segment data x(t) for the ith time and then decomposed by EMD; g ij (t) represents the jth intrinsic modal component in the key factor set obtained by EMD decomposition after adding white noise to the sequence segment data x(t) for the i-th time; Both i and j are non-zero natural numbers.
8. The regional power load forecasting method of the EEMD-LASSO-GM model according to claim 7 is characterized in that: The establishment and calculation steps of the GM model are as follows: S3.
1. Input and derivation of the original sequence, i.e., the key sequence g 0 (t) is used as the original sequence to accumulate once and derive the generated sequence g 1 (t); S3.2, construct differential equations, that is, according to the generated sequence g 1 (t) data to establish differential equations; S3.3, optimization of differential equations, i.e., optimizing the differential equations constructed in the previous step using the least squares method; S3.
4. Make a prediction, that is, based on the optimized differential equation, predict the change of power load in the area to be tested under given conditions.