Comprehensive energy load prediction method based on aggregation quadratic mode decomposition and Autoformer optimization

Abstract

The invention discloses a comprehensive energy load prediction method based on aggregation quadratic modal decomposition and Autoformer optimization, and the method comprises the steps: obtaining historical IES multivariate load data and climate, date rule and energy price factor data, and carrying out the preprocessing and feature engineering of the obtained data; iES multivariate load data are preliminarily decomposed through ICEEMDAN, aggregation and recombination are carried out on the IES multivariate load data through the FE value of each sub-component, and VMD decomposition is carried out on a random sequence with the maximum FE value to obtain different modal components; the hyper-parameter of the Autoformer prediction model is optimized through a BWO algorithm, so that the hyper-parameter of the Autoformer prediction model is optimal; and finally, respectively substituting different modal components into an Autoformer prediction model optimized by a Lion optimizer to respectively obtain prediction results of the sub-components, and respectively superposing and summing the prediction results of all the multivariate load sub-components to obtain final electric, cold and thermal load prediction results. According to the Lion optimization Autoformer prediction model based on aggregation quadratic mode decomposition provided by the invention, decomposition and prediction of IES load data are realized, certain robustness is enhanced, and the precision of IES load prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated energy system load forecasting, and relates to an integrated energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer. Background Art

[0002] With the advent of the new energy era, integrated energy systems (IES) are gaining increasing attention due to their importance in promoting energy structure transformation and sustainable development. Accurate multi-dimensional load forecasting is crucial to ensuring the reliable, cost-effective operation of IES. Multi-dimensional load forecasting not only helps optimize resource allocation and improve system efficiency, but also provides strong support for the operation of virtual power plants.

[0003] The current IES load forecasting method is insufficient in extracting feature information from IES load data with strong nonlinearity and non-stationarity. Substituting multivariate load data into the forecasting model affects the forecasting accuracy. At the same time, the traditional deep learning model has the phenomenon of information loss in long-sequence time series forecasting, resulting in low long-sequence time series forecasting accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer, which solves the problems existing in the prior art.

[0005] The technical solution adopted by the present invention is a comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer, which is implemented in the following steps:

[0006] S1: Obtain historical IES multivariate load data and climate, date rules, and energy price factor data;

[0007] S2: Perform visual analysis, missing value retrieval, and outlier check on the acquired data, and fill in missing values ​​and replace outliers through linear interpolation;

[0008] S3: Mark weekends / weekdays, holidays, months, and seasons respectively to obtain weekend / weekday, holiday, month, and seasonal features;

[0009] S4: Analyze all features by the maximum information coefficient, retain the features with the largest correlation, and discard the remaining features with the smallest correlation;

[0010] S5: Perform preliminary ICEEMDAN decomposition on the IES multivariate load data of electricity, cooling and heating, and decompose the non-stationary series into a finite number of IMF components with different frequency characteristics, recorded as IC_IMF;

[0011] S6: aggregate and reorganize IC_IMF by calculating the fuzzy entropy FE value of all IC_IMF;

[0012] S7: Perform VMD decomposition on the random sequence with the largest FE value in the recombined sequence based on the FE value to further reduce its complexity and reduce the difficulty of prediction. The subsequence obtained by aggregated secondary modal decomposition is recorded as Bv_IMF;

[0013] S8: Optimize the hyperparameters of the Autoformer prediction model using the BWO algorithm to achieve optimal hyperparameters.

[0014] S9: Input the Bv_IMF multivariate load subcomponents obtained by the aggregated secondary modal decomposition into the Autoformer prediction model optimized by the Lion optimizer to obtain the prediction results of the subcomponents respectively;

[0015] S10: The prediction results of all multi-component load sub-components are superimposed and summed up to obtain the final prediction results of electricity, cooling and heating loads.

[0016] In summary, this invention improves forecast accuracy by constructing multiple characteristic factors and employing aggregated quadratic modal decomposition (IES) load data to better leverage the characteristics of the forecast model. The Lion optimizer-optimized Autoformer forecasting model demonstrates high forecasting capabilities in the field of long-range time series (LSTF) forecasting. By incorporating different multivariate load subcomponents from aggregated quadratic modal decomposition into the forecasting model, superior forecast results are achieved. This method demonstrates robustness and improves the accuracy of IES load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] The present invention is a comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer. The detailed flow chart is as follows: Figure 1 As shown, the specific steps are as follows:

[0020] S1: Obtain historical IES multivariate load data and climate, date rules, and energy price factor data;

[0021] The sampling frequency of the acquired IES load data is once every hour.

[0022] S2: Perform visual analysis, missing value retrieval, and outlier inspection on the acquired data;

[0023] It is detected that the original data have different degrees of missing values ​​and outliers, and linear interpolation is used to fill the missing values ​​and replace the outliers.

[0024] S3: Mark holidays and seasons respectively to obtain holiday and seasonal features;

[0025] The seasonal factors and holiday factors hidden in the power load data are extracted, and the seasons are marked using the labeling method, with spring marked as 0, summer marked as 1, autumn marked as 2, and winter marked as 3; the holidays are also marked using the labeling method, with holidays marked as 0 and non-holidays marked as 1.

[0026] The most important factor in power load data is human behavior, but it is implicit in the data.

[0027] S4: Analyze the weather factor data by the maximum information coefficient, retain the features with the largest correlation, and discard the features with the smallest correlation;

[0028] The maximum information coefficient is used to calculate the correlation of weather factors, specifically:

[0029]

[0030] Where x and y are two sets of data whose correlation is required, and p(x,y) is the joint probability between variables x and y. The calculated correlation retains the highly correlated characteristic factors.

[0031] S5: Perform preliminary ICEEMDAN decomposition on the IES electricity, cooling and heating multi-load data, and decompose the non-stationary series into a finite number of IC_IMF components with different frequency characteristics;

[0032] ICEEMDAN is used to perform a modal decomposition on the multivariate load data, specifically:

[0033] S5.1: For the original load sequence x(n), add E1(w (i) )get:

[0034] x i =x+β0E1(w (i) )

[0035] S5.2: For x i Envelope calculation and averaging to obtain the first residual component r1 and the first modal component G IMF1 :

[0036] r1=<M(xi )>

[0037] G IMF1 =x-r1

[0038] S5.3: Calculate the second residual component r2, modal G IMF2 :

[0039] r2=<M(r1+β1E2(w (i) ))>

[0040] G IMF2 =r1-r2

[0041] S5.4: Repeat step 3 to calculate the kth residual component and mode:

[0042] r k =<M(r k-1 +β k-1 E k (w (i) ))>

[0043] G IMFk =r k-1 -r k

[0044] Among them E K () obtains the kth mode through the EMD algorithm; M() represents envelope calculation; < > represents mean value calculation; w (i) Indicates adding the i-th white noise, β k Expressed as the standard deviation of white noise.

[0045] S6: aggregate and reorganize IC_IMF by calculating the fuzzy entropy FE value of all IC_IMF;

[0046] The specific steps for calculating fuzzy entropy are:

[0047] S6.1: Assume that the one-dimensional time series x={x1,x2,...,x j ,x N}

[0048] S6.2: Embed the positive integers m and d into the time series for reconstruction. The vector is:

[0049]

[0050] in is the mean, m is the embedding dimension, and d is the time delay.

[0051] S6.3: Define similarity by fuzzy function and calculate its adjacent vectors Similarity

[0052] in, is the fuzzy membership function, yes and The maximum absolute value difference is , and r is the similarity tolerance.

[0053] S6.4: For each vector and its adjacent vectors, we have:

[0054]

[0055] S6.5: For time series Similarly, we can get:

[0056]

[0057] S6.6: Define fuzzy entropy F(m,d,n,r) as:

[0058]

[0059] For a finite time series, it can be approximated by statistics:

[0060]

[0061] S6.7: The subsequence aggregation method based on fuzzy entropy is as follows:

[0062]

[0063] Where K is the number of decompositions of IMFs; {F Ei |i=1,...,K} is the fuzzy entropy value of the i-th IMF.

[0064] S7.1: Perform VMD decomposition on the random sequence with the largest FE value among the recombinant sequences based on the FE value to further reduce its complexity and lower the difficulty of prediction;

[0065] Construct the variational constraint, which is expressed as:

[0066]

[0067] Where: K is the number of IMF subcomponents that need to be decomposed, {u k}、{ω k} respectively represent the Kth subcomponent and its center frequency after decomposition, is the Dirac function, and * is the convolution operator symbol.

[0068] Normalize all data:

[0069]

[0070] All data are mapped to [0, 1] through normalization, where x n Normalized data, x min is the minimum value in the original data, x max is the maximum value in the original data, x is a value in the original data.

[0071] S7.2: Solve the variational constraint expression in S7.1 and introduce the Lagrange penalty operator λ to transform the constrained variational problem into an unconstrained variational problem. The formula is:

[0072]

[0073] Where α represents the penalty factor.

[0074] S7.3: After step S7.2, the optimized modal subcomponents and center frequencies are searched for the saddle point of the Lagrange function, which is formulated as follows:

[0075]

[0076] In the formula Update using gradient descent.

[0077] S8: Optimize the hyperparameters of the Autoformer prediction model using the BWO algorithm to achieve optimal hyperparameters.

[0078] It is divided into three stages:

[0079] S8.1: Exploration phase. The position of the search agent is determined by the swimming of a pair of beluga whales, and its position is updated as follows:

[0080]

[0081] Where, t is the current iteration number, is the new position of the i-th beluga whale in the j-th dimension.

[0082] S8.2: Development phase. During the BWO development phase, the Levy flight strategy was introduced to capture prey. The mathematical model for capturing prey is expressed as:

[0083]

[0084] in and is the current position of the i-th beluga whale and a random beluga whale, is the new position of the i-th beluga whale, It is the best position for fish.

[0085] Lf The calculation method is as follows:

[0086]

[0087] where u and v are normally distributed random numbers, and β is a constant that defaults to 1.5.

[0088] S8.3: Whale fall stage. The position update model of the beluga whale is:

[0089]

[0090] The probability of a whale falling is W f is designed as a linear function:

[0091] W f =0.1-0.05t / T

[0092] The probability of a whale fall decreases from 0.1 in the initial iteration to 0.05 in the last iteration.

[0093] S9: Input the Bv_IMF multivariate load subcomponents obtained by the aggregated secondary modal decomposition into the Autoformer prediction model optimized by the Lion optimizer to obtain the prediction results of the subcomponents respectively;

[0094] Faced with massive, heterogeneous and high-dimensional IES load data, the Lion optimizer is used to optimize the Autoformer prediction model, resulting in a faster loss reduction rate.

[0095] S10: The prediction results of all the electricity, cooling and heating load sub-components are respectively superimposed and summed to obtain the final electricity, cooling and heating load prediction results.

[0096] The formula for calculating the mean square error between all true values ​​and predicted values ​​is as follows:

[0097]

[0098] Where y1 represents the actual value of the power load data, and y0 represents the predicted value of the power load data.

[0099] Implementation Examples

[0100] like Figure 1 , a comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer, including the following steps:

[0101] S1: Obtain historical IES multivariate load data and climate, date rules, and energy price factor data;

[0102] The ASU integrated energy system dataset from January 1, 2020 to December 31, 2022 is used, with a sampling time of once every hour.

[0103] S2: Perform visual analysis, missing value retrieval, and outlier check on the acquired data; fill in missing values ​​and replace outliers through linear interpolation.

[0104] S3: Mark holidays and seasons respectively to obtain holiday and seasonal features;

[0105] The seasons are marked using the labeling method, with spring marked as 0, summer marked as 1, autumn marked as 2, and winter marked as 3; the holidays are marked using the labeling method, with holidays marked as 0 and non-holidays marked as 1.

[0106] S4: Analyze the weather factor data by the maximum information coefficient, retain the features with the largest correlation, and discard the features with the smallest correlation;

[0107] The coupling between IES multiple loads significantly improves the prediction performance. The minimum temperature and average temperature are highly correlated with the multiple loads. The heating index and cooling index are most correlated with the cooling load. Energy prices have the greatest impact on the use of heat loads.

[0108] S5: Perform preliminary ICEEMDAN decomposition on the IES multivariate load data of electricity, cooling and heating, and decompose the non-stationary series into a finite number of IMF components with different frequency characteristics, recorded as IC_IMF;

[0109] S6: aggregate and reorganize IC_IMF by calculating the fuzzy entropy FE value of all IC_IMF;

[0110] S7: Perform VMD decomposition on the random sequence with the largest FE value in the recombined sequence based on the FE value to further reduce its complexity and reduce the difficulty of prediction. The subsequence obtained by aggregated secondary modal decomposition is recorded as Bv_IMF;

[0111] S8: Optimize the hyperparameters of the Autoformer prediction model using the BWO algorithm to achieve optimal hyperparameters.

[0112] S9: Input the Bv_IMF multivariate load subcomponents obtained by the aggregated secondary modal decomposition into the Autoformer prediction model optimized by the Lion optimizer to obtain the prediction results of the subcomponents respectively;

[0113] S10: The prediction results of all multi-component load sub-components are superimposed and summed up to obtain the final prediction results of electricity, cooling and heating loads.

[0114] The predicted results are evaluated as follows:

[0115]

[0116] The above content is only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer, characterized by: The following steps are involved: S1: Obtain historical IES multivariate load data and climate, date rules, and energy price factor data; S2: Perform visual analysis, missing value retrieval, and outlier check on the acquired data, and fill in missing values ​​and replace outliers through linear interpolation; S3: Mark weekends / weekdays, holidays, months, and seasons respectively to obtain weekend / weekday, holiday, month, and seasonal features; S4: Analyze all features by the maximum information coefficient, retain the features with the largest correlation, and discard the remaining features with the smallest correlation; S5: Perform preliminary ICEEMDAN decomposition on the IES multivariate load data of electricity, cooling and heating, and decompose the non-stationary series into a finite number of IMF components with different frequency characteristics, recorded as IC_IMF; S6: aggregate and reorganize IC_IMF by calculating the fuzzy entropy FE value of all IC_IMF; S7: Perform VMD decomposition on the random sequence with the largest FE value in the recombined sequence based on the FE value to further reduce its complexity and reduce the difficulty of prediction. The subsequence obtained by aggregated secondary modal decomposition is recorded as Bv_IMF; S8: Optimize the hyperparameters of the Autoformer prediction model using the BWO algorithm to achieve optimal hyperparameters. S9: Input the Bv_IMF multivariate load subcomponents obtained by the aggregated secondary modal decomposition into the Autoformer prediction model optimized by the Lion optimizer to obtain the prediction results of the subcomponents respectively; S10: The prediction results of all multi-component load sub-components are superimposed and summed up to obtain the final prediction results of electricity, cooling and heating loads.

2. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: The climate and energy price factor data obtained in step S1 include wind speed, air pressure, temperature, relative humidity, wind direction, precipitable water, surface albedo, dew point, electricity price, natural gas futures, oil futures, and coal futures.

3. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: In step S2, the original data contains missing values ​​and outliers to varying degrees, and linear interpolation is used to fill in missing values ​​and replace outliers.

4. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized autoformer according to claim 1 is characterized by: In step S3, the weekends / weekdays, holidays, months and seasons implied in the power load data need to be extracted.

5. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 4 is characterized by: In step S3, the seasons are marked by a marking method, with Spring Festival marked as 0, summer marked as 1, autumn marked as 2, and winter marked as 3; holidays are marked by a marking method, with holidays marked as 0 and non-holidays marked as 1.

6. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: In step S4, the maximum information coefficient is used to calculate the correlation of weather factors, specifically: Where x and y are two sets of data whose correlation is required, and p(x,y) is the joint probability between variables x and y. The calculated correlation retains the highly correlated characteristic factors.

7. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: In step S5, the IES electrical, cooling, and heating multi-load data are preliminarily decomposed using ICEEMDAN to obtain a series of modal components with different frequencies.

8. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: Step S8 includes: S8.1: Exploration phase. The location of the search agent is determined by the swimming of a pair of beluga whales; S8.2: Development phase. During the BWO development phase, the Levy flight strategy was introduced to capture prey. S8.3: Whale fall phase. Update model of the beluga whale's position.

9. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: In step S9, the Bv_IMF multivariate load subcomponents obtained by the aggregated secondary modal decomposition are input into the Autoformer prediction model optimized by the Lion optimizer to obtain the prediction results of the subcomponents respectively. Faced with massive, heterogeneous, and high-dimensional IES load data, the Lion optimizer was used to optimize the Autoformer prediction model, resulting in a faster rate of loss reduction.

10. The comprehensive energy load forecasting method based on aggregated quadratic modal decomposition and optimized Autoformer according to claim 1 is characterized by: Step S10 obtains the final electricity, cooling and heating load prediction results by superimposing and summing the prediction results of all multi-component load sub-components.

Images