Short-term load prediction method considering error compensation and mode decomposition

Abstract

The invention discloses a short-term load prediction method considering error compensation and modal decomposition. The short-term load prediction method comprises the following steps: S1, acquiring load data and characteristic data of a power system; s2, carrying out the preliminary decomposition of the load data of the power system based on ICEEMDAN, and obtaining an intrinsic mode function; s3, training a first network model TCN-GRU by taking the intrinsic mode function and the feature data as input to realize initial short-term load initial prediction; s4, comparing the preliminary short-term load prediction result with a target load, constructing an error sequence, and reducing the unpredictability of the error sequence by adopting AVMD to obtain a variational mode function; and S5, training the second network model TCN-GRU by taking the variational mode function as input, and correcting the initial short-term load initial prediction result to obtain a final short-term load prediction result. The problems that in the prior art, a single prediction model is insufficient in precision, and the effect of an error sequence on the aspect of improving the load prediction precision is not fully utilized are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, in particular to power load forecasting technology, and specifically to a TCN-GRU short-term load forecasting method considering error compensation and modal decomposition. Background Art

[0002] Load forecasting is an important means to maintain the balance of power supply and demand in the power grid and is also the basis for the operation of the power market. Power system planning and operation rely on accurate load forecasting in different time frames. The accuracy of load forecasting has an important impact on the stability and operation cost of the power grid. Current load forecasting work mainly focuses on developing effective feature extraction methods and advanced forecasting models, lacks the application of combined models and ignores the role of load forecasting errors in improving forecasting accuracy.

[0003] Patent application number 202311496873.1 proposes the use of a single LSTM network model for power load forecasting, but the LSTM structure is complex, has many training parameters, and has limited ability to extract time series data. Patent application number 202410013103.5 uses GCN-CBAM-BiGRU as a combined forecasting model and effectively reduces the forecast error by optimizing the model parameters through Bayesian optimization.

[0004] In summary, the present invention combines signal decomposition theory and deep learning model to propose a short-term load forecasting method and system considering error compensation and modal decomposition. The complexity of the original power load sequence and error sequence are reduced respectively through effective signal decomposition technology, and are input into the TCN-GRU combined model to realize load forecasting and error correction. Summary of the invention

[0005] The purpose of the present invention is to propose a short-term load forecasting method and system considering error compensation and modal decomposition, which mainly combines improved complete set empirical mode decomposition, time convolutional network, gated recurrent unit and adaptive variational mode decomposition to achieve full extraction of load data information and full utilization of error sequence, thereby improving the accuracy of power load forecasting.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A short-term load forecasting method considering error compensation and modal decomposition includes the following steps:

[0008] Step S1: Collect power system load data D 1 and the corresponding characteristic data D 2 ;

[0009] Step S2: Based on the improved complete set empirical mode decomposition, the load data D1 Decompose and obtain a set of intrinsic mode functions IMFs that characterize load characteristics;

[0010] Step S3: Using the intrinsic mode function IMF obtained in step S2 and the characteristic data D obtained in step S1 2 As input, train the first network model TCN-GRU to obtain the preliminary short-term load forecast result Y t ;

[0011] Step S4: Using the preliminary short-term load forecast result Y t Compare it with the corresponding target load in the load data D1 to construct the short-term load forecast error e t , using AVMD to estimate the short-term load forecast error e t Decompose and obtain a set of variational mode functions VMF;

[0012] Step S5: Taking the variational mode function VMF as input, the second network model TCN-GRU is used as the error prediction model to obtain the error prediction result E t , the preliminary short-term load forecast result Y of step S3 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

[0013] Furthermore, in step S1, the power system load data D 1 The sampling is performed at 0.5h intervals, and contains 17520 sampling points in total. 2 Contains characteristic information such as temperature, humidity, rainfall and electricity price.

[0014] Furthermore, in step S2, the specific steps of using the improved complete set empirical mode decomposition are as follows:

[0015] Step 1: In the power system load data D 1 Extract the load sequence from t , according to the formula y n,t =y t +β 0 E 1 (ω) Construct a new decomposition sequence y n ,t:

[0016] Step 2: Calculate the envelope res of the new decomposition sequence yn,t 1 = <M(y n,t )>, get the first residual component res 1 :

[0017] Step 3: Use the original load sequence y t Subtract res1 Get the first intrinsic mode function IMF 1 ,

[0018] Step 4: Based on the IMF 2 =res 1 - <M(r 1 +βE 1 (ω))>Calculate the second intrinsic mode function IMF2;

[0019] Step 5: According to the IMF formula k =res k - <M(res k-1 +βE k (ω))>Calculate the kth intrinsic mode function IMF k ;

[0020] Step 6: Combine IMF 1 ~IMF k , complete the load data D 1 decomposition;

[0021] Among them, y t is the original load sequence; y n,t is the new decomposition sequence; <·> is the average value calculation; M(·) is the envelope calculation; E k (·) is the kth modal component of EMD decomposition; res k is the kth residual component; IMF k is the kth intrinsic mode function; x is the original sequence; ω is Gaussian white noise; β is the noise coefficient.

[0022] Further, in step S3, the details are as follows:

[0023] Step 1: Combine the intrinsic mode function IMF and the power system load data D 1 Add to feature data D 2 In this paper, we construct a new feature dataset D 3 ;

[0024] Step 2: Convert the new feature dataset D to 3 Divide into training set, test set and validation set;

[0025] Step 3: Train the first network model TCN-GRU using the training set data;

[0026] Step 4: Perform prediction based on the trained first network model TCN-GRU to obtain the preliminary short-term load forecast result Y t .

[0027] Furthermore, in step S4, the following steps are specifically included:

[0028] Step 1: Compare the preliminary short-term load forecast results Y t and D 3 The actual load in the short-term load forecast error e is obtained t ;

[0029] Step 2: According to the minimum K value, take the short-term load forecast error e t Constructing the variational model for the input is as follows:

[0030]

[0031] In the formula, e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω k is the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, and * is the convolution operation.

[0032] Step 3: Based on the quadratic penalty factor and Lagrange multiplier, the variational model is transformed to obtain the Lagrangian function. The results are as follows:

[0033]

[0034] Where L is the Lagrangian function; α is the quadratic penalty factor; λ is the Lagrangian multiplier; e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω k is the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, and * is the convolution operation.

[0035] Step 4: Optimize the Lagrangian function by alternating direction multiplier method and iteratively update u k ,ω k and λ until the following equation is satisfied:

[0036]

[0037] In the formula, ε is the discriminant parameter, n is the number of iterations, for The corresponding Fourier transform.

[0038] Step 5: Define the loss coefficient Set the threshold μ. When e<μ is satisfied, stop the iteration, implement VMD decomposition based on the current modal decomposition number K, and obtain the variational mode function VMF; otherwise, add 1 to the K value and repeat steps 2 to 5.

[0039] Further, in step S5, the details are as follows:

[0040] Step 1: Divide the variational mode function VMF into training set, test set and validation set according to a fixed ratio;

[0041] Step 2: Train the second network model TCN-GRU using the training set data;

[0042] Step 3: Make predictions based on the trained second network model TCN-GRU to obtain the error prediction result E t ;

[0043] Step 4: Convert the initial short-term load forecast result Y of weight 4 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

[0044] Compared with the prior art, the present invention has the following technical effects:

[0045] 1) The present invention utilizes signal decomposition and deep learning technology to decompose complex power load data, and considers the role of load forecasting error in improving load forecasting accuracy, thereby further improving forecasting accuracy;

[0046] 2) The present invention decomposes the power load into intrinsic mode functions (IMFs) of different frequencies by using an improved complete set empirical mode decomposition, thereby expanding the input feature data set;

[0047] 3) The present invention combines the advantages of TCN feature extraction and GRU nonlinear fitting to construct a combined prediction model to achieve complementary advantages;

[0048] 4) The present invention adopts AVMD to find the optimal decomposition number of the error sequence, and can complete the effective decomposition of VMD no matter how complex the sequence to be decomposed is. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0050] Figure 1 It is a prediction structure diagram of the present invention;

[0051] Figure 2 It is a system structure diagram of the present invention;

[0052] Figure 3 A comparison chart of the first quarter results predicted by the present invention;

[0053] Figure 4 A comparison chart of the predicted second quarter results for the present invention;

[0054] Figure 5 A comparison chart of the third quarter results predicted by the present invention;

[0055] Figure 6 This is a comparison chart of the fourth quarter results predicted by the present invention. DETAILED DESCRIPTION

[0056] like Figure 1 As shown, a short-term load forecasting method considering error compensation and modal decomposition includes the following steps:

[0057] Step S1: Collect power system load data D 1 and the corresponding characteristic data D 2 ;

[0058] Step S2: Based on the improved complete set empirical mode decomposition, the load data D 1 Decompose and obtain a set of intrinsic mode functions IMFs that characterize load characteristics;

[0059] Step S3: Using the intrinsic mode function IMF obtained in step S2 and the characteristic data D obtained in step S1 2 As input, train the first network model TCN-GRU to obtain the preliminary short-term load forecast result Y t ;

[0060] Step S4: Using the preliminary short-term load forecast result Y t Compare it with the corresponding target load in the load data D1 to construct the short-term load forecast error e t , using AVMD to estimate the short-term load forecast error e t Decompose and obtain a set of variational mode functions VMF;

[0061] Step S5: Taking the variational mode function VMF as input, the second network model TCN-GRU is used as the error prediction model to obtain the error prediction result E t , the preliminary short-term load forecast result Y of step S3 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

[0062] Furthermore, in step S1, the power system load data D 1 The sampling is performed at 0.5h intervals, and contains 17520 sampling points in total. 2 Contains characteristic information such as temperature, humidity, rainfall and electricity price.

[0063] Furthermore, in step S2, the specific steps of using the improved complete set empirical mode decomposition are as follows:

[0064] Step 1: In the power system load data D 1 Extract the load sequence from t , according to the formula y n,t =y t +β 0 E 1 (ω) Construct a new decomposition sequence y n,t :

[0065] Step 2: Calculate the envelope res of the new decomposition sequence yn,t 1 = <M(y n,t )>, get the first residual component res 1 :

[0066] Step 3: Use the original load sequence y t Subtract res 1 Get the first intrinsic mode function IMF 1 ,

[0067] Step 4: Based on the IMF 2 =res 1 - <M(r 1 +βE 1 (ω))>Calculate the second intrinsic mode function IMF2;

[0068] Step 5: According to the IMF formula k =res k - <M(res k-1 +βE k (ω))>Calculate the kth intrinsic mode function IMFk;

[0069] Step 6: Combine IMF 1 ~IMF k , complete the load data D 1 decomposition;

[0070] Among them, y t is the original load sequence; y n,t is the new decomposition sequence; <·> is the average value calculation; M(·) is the envelope calculation; E k (·) is the kth modal component of EMD decomposition; res k is the kth residual component; IMF k is the kth intrinsic mode function; x is the original sequence; ω is Gaussian white noise; β is the noise coefficient.

[0071] Further, in step S3, the details are as follows:

[0072] Step 1: Combine the intrinsic mode function IMF and the power system load data D 1 Add to feature data D 2 In this paper, we construct a new feature dataset D 3 ;

[0073] Step 2: Convert the new feature dataset D to 3 Divide into training set, test set and validation set;

[0074] Step 3: Train the first network model TCN-GRU using the training set data;

[0075] Step 4: Perform prediction based on the trained first network model TCN-GRU to obtain the preliminary short-term load forecast result Y t .

[0076] Furthermore, in step S4, the following steps are specifically included:

[0077] Step 1: Compare the preliminary short-term load forecast results Y t and D 3 The actual load in the short-term load forecast error e is obtained t ;

[0078] Step 2: According to the minimum K value, take the short-term load forecast error e t Constructing the variational model for the input is as follows:

[0079]

[0080] In the formula, e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω k is the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, and * is the convolution operation.

[0081] Step 3: Based on the quadratic penalty factor and Lagrange multiplier, the variational model is transformed to obtain the Lagrangian function. The results are as follows:

[0082]

[0083] Where L is the Lagrangian function; α is the quadratic penalty factor; λ is the Lagrangian multiplier; e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω kis the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, and * is the convolution operation.

[0084] Step 4: Optimize the Lagrangian function by alternating direction multiplier method and iteratively update u k ,ω k and λ until the following equation is satisfied:

[0085]

[0086] In the formula, ε is the discriminant parameter, n is the number of iterations, for The corresponding Fourier transform.

[0087] Step 5: Define the loss coefficient Set the threshold μ. When e<μ is satisfied, stop the iteration, implement VMD decomposition based on the current modal decomposition number K, and obtain the variational mode function VMF; otherwise, add 1 to the K value and repeat steps 2 to 5.

[0088] Further, in step S5, the details are as follows:

[0089] Step 1: Divide the variational mode function VMF into training set, test set and validation set according to a fixed ratio;

[0090] Step 2: Train the second network model TCN-GRU using the training set data;

[0091] Step 3: Make predictions based on the trained second network model TCN-GRU to obtain the error prediction result E t ;

[0092] Step 4: Convert the initial short-term load forecast result Y of weight 4 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

[0093] Verification test: The annual load is divided into four quarters according to time, and the root mean square error, mean absolute error and mean absolute percentage error are used as evaluation indicators, as follows:

[0094]

[0095] Where RMSE is the root mean square error; MAE is the mean absolute error; MAPE is the mean absolute percentage error; N is the sample size; Y is the t act is the actual value; Y t pre is the predicted value.

[0096] The prediction results are as follows Figure 3 to Figure 6 As shown in the figure, the prediction results obtained by training the model under different input data are close to the true value, with only a small number of data points having relatively large deviations. The overall prediction error is low and the prediction performance is good.

[0097] Example 2

[0098] The present invention provides a short-term load forecasting system considering error compensation and modal decomposition, such as Figure 2 As shown, including:

[0099] Data acquisition module: used to obtain power load data and characteristic data;

[0100] Load decomposition module: used to reduce the complexity of power load data and decompose the power load into intrinsic mode functions of different frequencies;

[0101] Load forecasting module: used to input the intrinsic mode function and the corresponding characteristic data into the first network model TCN-GRU model to obtain the load forecasting result;

[0102] The error acquisition module is used to compare the predicted load with the actual load to obtain the load error;

[0103] The error decomposition module is used to reduce the unpredictability of the fixed-length error sequence and obtain the error subsequence;

[0104] The error compensation module is used to input the error subsequence into the second network model TCN-GRU model, complete the prediction of the error sequence, and correct the load prediction result.

Claims

1. A short-term load forecasting method considering error compensation and modal decomposition, characterized in that: The following steps are involved: Step S1: Collect power system load data D1 and corresponding characteristic data D2; Step S2: Decompose the load data D1 based on the improved complete set empirical mode decomposition to obtain a set of intrinsic mode functions IMFs that characterize the load characteristics; Step S3: Using the intrinsic mode function IMF obtained in step S2 and the characteristic data D2 obtained in step S1 as input, train the first network model TCN-GRU to obtain the preliminary short-term load forecasting result Y t ; Step S4: Using the preliminary short-term load forecast result Y t Compare it with the corresponding target load in the load data D1 to construct the short-term load forecast error e t , using AVMD to estimate the short-term load forecast error e t Decompose and obtain a set of variational mode functions VMF; Step S5: Taking the variational mode function VMF as input, the second network model TCN-GRU is used as the error prediction model to obtain the error prediction result E t , the preliminary short-term load forecast result Y of step S3 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

2. The method according to claim 1, characterized in that In step S1, the power system load data D1 is sampled at a certain time interval, and includes a total of several sampling points; the corresponding characteristic data D2 includes characteristic information such as temperature, humidity, rainfall and electricity price.

3. The method according to claim 1, characterized in that: In step S2, the following steps are specifically included: Step 2-1: Extract the load sequence from the power system load data D1, denoted as y t , according to the formula y n,t =y t +βE1(ω) constructs a new decomposition sequence y n,t : Step 2-2: Calculate the envelope of the new decomposition sequence yn,t res1= <M(y n,t )>, get the first residual component res1: Step 2-3: Use the original load sequence y t Subtract res1 to get the first intrinsic mode function IMF1, Step 2-4: According to IMF2=res1-<M(res1+βE2(ω))> Calculate the second intrinsic mode function IMF2; Step 2-5: According to IMF k =res k-1 - <M(res k-1 +βE k (ω))>Calculate the kth intrinsic mode function IMF k ; Step 2-6: Combine IMF1 to IMF k , complete the decomposition of load data D1; Among them, y t is the original load sequence; y n,t is the new decomposition sequence; <·> is the average value calculation; M(·) is the envelope calculation; E k (·) is the kth modal component of EMD decomposition; res k is the kth residual component; IMF k is the kth intrinsic mode function; x is the original sequence; ω is Gaussian white noise; β is the noise coefficient.

4. The method according to claim 1, characterized in that: In step S3, the following steps are specifically included: Step 3-1: Add the intrinsic mode function IMF and the power system load data D1 to the characteristic data D2 to construct a new characteristic data set D3; Step 3-2: Divide the new feature dataset D3 into training set, test set and validation set according to a fixed ratio; Step 3-3: Train the first network model TCN-GRU using the training set data; Step 3-4: Perform prediction based on the trained first network model TCN-GRU to obtain the preliminary short-term load forecast result Y t .

5. The method according to claim 1, characterized in that In step S4, the following steps are specifically included: Step 4-1: Compare the preliminary short-term load forecast results Y t and the actual load in D3, the short-term load forecast error e is obtained t ; Step 4-2: According to the minimum K value, take the short-term load forecast error e t Constructing the variational model for the input is as follows: In the formula, e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω k is the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, * is the convolution operation; Step 4-3: Based on the quadratic penalty factor and Lagrange multiplier, the variational model is transformed to obtain the Lagrangian function. The results are as follows: Where L is the Lagrangian function; α is the quadratic penalty factor; λ is the Lagrangian multiplier; e t is the short-term load forecast error; is the partial derivative operator; j is the imaginary unit; u k (t) is the kth subsequence; ω k is the center frequency of the kth subsequence; t is time; K is the modal decomposition number; δ(t) is the Dirac function, * is the convolution operation; Step 4-4: Optimize the Lagrangian function by alternating direction multiplier method and iteratively update u k ,ω k and λ until the following equation is satisfied: In the formula, ε is the discriminant parameter, n is the number of iterations, for The corresponding Fourier transform; Step 4-5: Define loss coefficients Set the threshold μ. When e<μ is satisfied, stop the iteration, implement VMD decomposition based on the current modal decomposition number K, and obtain the variational mode function VMF; otherwise, add 1 to the K value and repeat steps 4-2 to 4-5.

6. The method according to claim 1, characterized in that In step S5, the details are as follows: Step 5-1: Divide the variational mode function VMF into a training set, a test set, and a validation set according to a fixed ratio; Step 5-2: Train the second network model TCN-GRU using the training set data; Step 5-3: Make predictions based on the trained second network model TCN-GRU to obtain the error prediction result E t ; Step 5-4: Substitute the initial short-term load forecast result Y of weight 4 t And the error prediction result E t Superposition, get the final short-term load forecast result Y L .

7. A short-term load forecasting system considering error compensation and modal decomposition, characterized in that: include: Data acquisition module: used to obtain power load data and characteristic data; Load decomposition module: used to reduce the complexity of power load data and decompose the power load into intrinsic mode functions of different frequencies; Load forecasting module: used to input the intrinsic mode function and the corresponding characteristic data into the first network model TCN-GRU model to obtain the load forecasting result; The error acquisition module is used to compare the predicted load with the actual load to obtain the load error; The error decomposition module is used to reduce the unpredictability of the fixed-length error sequence and obtain the error subsequence; The error compensation module is used to input the error subsequence into the second network model TCN-GRU model, complete the prediction of the error sequence, and correct the load prediction result.

Images