Hybrid model load prediction method based on improved feature processing

Abstract

The invention discloses a hybrid model load prediction method based on improved feature processing. The method comprises the following steps: firstly, filtering high-frequency noise in load data by using an improved filtering algorithm based on an SCSO algorithm; using a periodic feature mapping method to obtain periodic features of the load data, and forming a training set and a test set with other features; finally, building an SCSO-MHA-GRU hybrid prediction model, introducing a multi-head self-attention layer MHA into a GRU network, improving the capability of capturing the relation among multiple features, and optimizing the learning rate by using an SCSO algorithm; and loading data required to be predicted into the trained SCSO-MHA-GRU hybrid model to predict a load value at a future moment. According to the method, for short-term load prediction of the low-voltage resident area, the load value of the next one hour is predicted by adopting the improved feature processing hybrid model, and experiments show that the multi-feature load data prediction precision of the low-voltage resident area is effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of load forecasting in electric power systems, and in particular relates to a hybrid model residential area load forecasting method based on improved feature processing. Background Art

[0002] Sand Cat Swarm Optimization (SCSO) is a new metaheuristic algorithm proposed by Amir Seyyedabbasi and Farzad Kiani in 2022. Sand Cat Swarm Optimization simulates the two main behaviors of sand cats: searching for prey and attacking prey.

[0003] The core idea of ​​the Multi-head Attention (MHA) mechanism is to split the input sequence into multiple subsequences, then perform independent attention calculations on each subsequence, and finally concatenate the outputs of these subsequences to obtain the final output. This allows the model to learn different feature representations in different attention directions, thereby improving the generalization and expression capabilities of the model.

[0004] Gated Recurrent Unit (GRU) is a commonly used gated recurrent neural network. It uses update gate to replace the input gate and forget gate in LSTM, and introduces the concept of reset gate, so as to achieve similar effects as LSTM network with fewer parameters and smaller computational overhead, and the training speed is faster. Summary of the invention

[0005] Purpose of the invention: The present invention proposes a hybrid model load forecasting method based on improved feature processing, which improves the accuracy of multi-feature load data forecasting in residential areas.

[0006] Technical solution: The present invention proposes a hybrid model load forecasting method based on improved feature processing, comprising the following steps:

[0007] (1) A new cycle feature mapping method proposed in the present invention is used to digitize the monthly cycle and hourly cycle to replace the seasonal characteristics, and the distance correlation coefficient is used to screen the features to mine the influencing factors with strong correlation with the load data;

[0008] (2) Using a high-frequency filtering algorithm based on SCSO optimization, the high-frequency components in the load data are filtered out to obtain filtered load data;

[0009] (3) Build an SMG hybrid prediction model, introduce a multi-head self-attention mechanism into the model to capture multiple features, and optimize the model learning rate through the SCSO algorithm. Then, the feature data obtained by filtering in step 1 and the filtered load data obtained in step 2 are normalized and input into the model for prediction;

[0010] The step (1) comprises the following steps:

[0011] (11) First, calculate the eigenvalue S after the monthly cycle mapping of the load data cast , the formula is as follows:

[0012]

[0013] Where: period m is the periodic value of the monthly characteristic of the load data, and its value is 12; m is the month of each year, and its value range is [1, 12]; A is the scaling factor, and its value is 10.

[0014] (12) Then calculate the characteristic value D after the load data hourly cycle mapping cast , the formula is as follows:

[0015]

[0016] Where: period d is the periodic value of the hourly characteristic of load data, and its value is 24; d is the number of hours per day, and its value range is [0, 23].

[0017] (13) Finally, the two eigenvalues ​​are superimposed to obtain the periodic eigenvalue SD after mapping. cast , the formula is as follows:

[0018] SD cast =S cast +D cast

[0019] The step (2) comprises the following steps:

[0020] (21) Take out the load data and record it as x(n);

[0021] (22) Calculate the frequency axis f of the load data k , that is, where N is the length of the time series data, the formula is as follows:

[0022]

[0023] Where: j is the index of the sequence, N is the length of the sequence, t1 and t0 are the first and second time points of the time series data respectively. In order to facilitate the subsequent algorithm optimization, the frequency axis is enlarged 1000 times.

[0024] (23) Perform fast Fourier transform (FFT) on the load data to obtain the frequency domain signal X k , the formula is as follows:

[0025]

[0026] Where: x n The time domain signal representing the load data, X k represents the frequency domain signal of the load data, N represents the length of the signal, and k represents the subscript of the frequency domain signal of the load data.

[0027] (24) Initialize the filter threshold T, and then calculate the R between the predicted value and the original load 2 The formula is as follows, and this R 2 The value is used as the fitness function of the SCSO algorithm to participate in the optimization and obtain the optimal threshold T best .

[0028]

[0029] Where: x represents the original load value, represents the actual load average value, Indicates the load value after filtering at each moment.

[0030] (25) All frequencies with absolute values ​​greater than the optimal threshold T best The load data is set to zero, and the formula is as follows:

[0031] X k =0,|f k |>T best

[0032] (26) Perform inverse Fourier transform (IFFT) on the filtered frequency domain load data to obtain the filtered time domain load data The formula is as follows:

[0033] The step (3) comprises the following steps:

[0034] (31) Using the periodic feature SD obtained by screening cast , temperature T, solar radiation SR and filtered load data Composition of the dataset Then the dataset is normalized to MIN-MAX to eliminate the impact of different dimensions on deep learning training.

[0035] (32) A GRU network was built and an MHA layer was added to the GRU network to capture the complex relationship between multiple input features and extract the long-term dependencies in the load data. The SCSO algorithm was then used to optimize the learning rate of deep learning.

[0036] (33) Using the SMG hybrid forecasting model with optimal parameters, the load forecasting result x is obtained pred .

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. The periodic feature SD obtained by the periodic feature mapping method proposed in the present invention cast Compared with the traditional season mapping method, the correlation between the seasonal characteristics mapped and the load data is significantly improved, and the distance correlation coefficient reaches 0.65.

[0039] 2. The high-frequency filtering method based on SCSO optimization proposed in the present invention not only effectively reduces the noise of the original load data, but also does not cause a significant loss in prediction accuracy, thereby effectively improving the model's fitting ability and prediction accuracy.

[0040] 3. The SMG hybrid prediction model built by the present invention improves the ability of the model to capture the relationship and long-term dependency between multiple input features by introducing the MHA layer in the GRU network. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the present invention;

[0042] Figure 2 The heat map of distance correlation coefficients calculated for the seven features and load data;

[0043] Figure 3 It is a comparison chart between the filtered load data and the original load data;

[0044] Figure 4 It is a comparison chart of the fitting degree between the predicted data of six models and the real load data; DETAILED DESCRIPTION

[0045] The present invention is further described in detail below with reference to the accompanying drawings and tables.

[0046] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0047] The data used in this invention are the real power load data and related climate data of a certain area in Jiangsu. The time range is from January 1, 2022 to December 31, 2022, the sampling interval is 1 hour, and there are a total of 8759 load values ​​at each moment. Then, LSTM, GRU, TCN, MHA-GRU, and VMD-GRU models are built to compare the prediction results. The model training results show that the comparison model does not perform well in the field of residential area load prediction. Therefore, this invention proposes a hybrid model load prediction method based on improved feature processing. The specific operation steps are as follows:

[0048] 1. A new cycle feature mapping method proposed in the present invention is used to digitize the monthly cycle and hourly cycle to replace the seasonal characteristics, and the distance correlation coefficient is used to perform feature screening to mine the influencing factors with strong correlation with the load data.

[0049] 1) First, calculate the eigenvalue S of the load data after monthly cycle mapping cast , the formula is as follows:

[0050]

[0051] Where: period m is the periodic value of the monthly characteristic of the load data, and its value is 12; m is the month of each year, and its value range is [1, 12]; A is the scaling factor, and its value is 10.

[0052] 2) Then calculate the characteristic value D after the load data hourly cycle mapping cast , the formula is as follows:

[0053]

[0054] Where: period d is the periodic value of the hourly characteristic of load data, and its value is 24; d is the number of hours per day, and its value range is [0, 23].

[0055] 3) Finally, the two eigenvalues ​​are superimposed to obtain the periodic eigenvalue SD after mapping cast , the formula is as follows:

[0056] SD cast =S cast +D cast

[0057] 2. Use the high-frequency filtering algorithm based on SCSO optimization to filter out the high-frequency components in the load data and obtain the filtered load data.

[0058] 1) Take out the load data and record it as x(n);

[0059] 2) Calculate the frequency axis f of the load data k , that is, where N is the length of the time series data, the formula is as follows:

[0060]

[0061] Where: j is the index of the sequence, N is the length of the sequence, t1 and t0 are the first and second time points of the time series data respectively. In order to facilitate the subsequent algorithm optimization, the frequency axis is enlarged 1000 times.

[0062] 3) Perform fast Fourier transform (FFT) on the load data to obtain the frequency domain signal X k , the formula is as follows:

[0063]

[0064] Where: x n The time domain signal representing the load data, X k represents the frequency domain signal of the load data, N represents the length of the signal, and k represents the subscript of the frequency domain signal of the load data.

[0065] 4) Initialize the filter threshold T, and then calculate the R of the predicted value and the original load 2 The formula is as follows, and this R 2 The value is used as the fitness function of the SCSO algorithm to participate in the optimization and obtain the optimal threshold T best .

[0066]

[0067] Where: x represents the original load value, represents the actual load average value, Indicates the load value after filtering at each moment.

[0068] 5) Set all frequencies with absolute values ​​greater than the optimal threshold T best The load data is set to zero, and the formula is as follows:

[0069] X k =0,|f k |>T best

[0070] 6) Perform inverse Fourier transform (IFFT) on the filtered frequency domain load data to obtain the filtered time domain load data The formula is as follows:

[0071]

[0072] 3. Build an SMG hybrid prediction model. Introduce a multi-head self-attention mechanism in the model to capture multiple features, and optimize the model learning rate through the SCSO algorithm. Then, normalize the feature data obtained by filtering in step 1 and the filtered load data obtained in step 2 and input them into the model for prediction.

[0073] 1) Using the periodic feature mapping method proposed in the present invention to obtain the periodic feature SD cast , used to expand feature data.

[0074] 2) Apply the high-frequency filtering algorithm based on the SCSO algorithm proposed in the present invention to filter the original load data to obtain the filtered load data

[0075] 3) Calculate the distance correlation coefficient for feature screening. Use the periodic feature SD obtained by screening cast , temperature T, solar radiation SR and filtered load data Composition of the dataset Then the dataset is normalized to MIN-MAX to eliminate the impact of different dimensions on deep learning training.

[0076] 4) Build a GRU network and add an MHA layer to the GRU network to capture the complex relationship between multiple input features and extract long-term dependencies in load data. Then use the SCSO algorithm to optimize the learning rate of deep learning.

[0077] 5) Use the SMG hybrid forecasting model with optimal parameters to obtain the load forecast result x pred The present invention builds LSTM, GRU, TCN, MHA-GRU, and VMD-GRU models respectively to compare the prediction results. Figure 4 As shown in the figure, it can be seen that the model of the present invention has a higher degree of fit with the actual load data than other models. The present invention uses mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) and precision (R 2 ×100) was used as the evaluation index for quantitative analysis, as shown in the following table.

[0078]

[0079] It can be seen from the table that when the data is not processed, the model is difficult to fit, resulting in low accuracy. Even if a model with stronger prediction ability (such as TCN, MHA-GRU) is replaced, it is still impossible to produce a good fit for the overly sharp residential area load data. Although the prediction accuracy can be effectively improved by decomposing and re-forecasting, the loss of accuracy is large. The model proposed in the present invention can find an optimal balance between loss of accuracy and improvement of fitting ability. Its MAPE index is 0.0211% lower than that of the VMD-GRU model and 0.0484% lower than that of the MHA-GRU model. The model proposed in the present invention is significantly better than other models in the field of short-term prediction of residential area load, with an accuracy of 99.2296%.

Claims

1. A hybrid model load forecasting method based on improved feature processing, characterized in that: The following steps are involved: (1) A new cycle feature mapping method proposed in the present invention is used to digitize the monthly cycle and hourly cycle to replace the seasonal characteristics, and the distance correlation coefficient is used to screen the features to mine the influencing factors with strong correlation with the load data; (2) Using a high-frequency filtering algorithm based on SCSO optimization, the high-frequency components in the load data are filtered out to obtain filtered load data; (3) An SMG hybrid prediction model was built. A multi-head self-attention mechanism was introduced into the model to capture multiple features. The model learning rate was optimized through the SCSO algorithm. The feature data obtained by filtering in step 1 and the filtered load data obtained in step 2 were normalized and input into the model for prediction.

2. A hybrid model load forecasting method based on improved feature processing according to claim 1, characterized in that: The step (1) comprises the following steps: (11) Calculate the eigenvalue S after monthly cycle mapping of load data cast , the formula is as follows: Where: period m is the periodic value of the monthly characteristic of the load data, and its value is 12; m is the month of each year, and its value range is [1, 12]; A is the scaling factor, and its value is 10. (12) Calculate the characteristic value D after the load data is mapped to the hourly cycle cast , the formula is as follows: Where: period d is the periodic value of the hourly characteristic of load data, and its value is 24; d is the number of hours per day, and its value range is [0, 23]. (13) The two eigenvalues ​​are superimposed to obtain the periodic eigenvalue SD after mapping cast , the formula is as follows: SD cast =S cast +D cast 。 3. The hybrid model load forecasting method based on improved feature processing according to claim 1 is characterized in that: The steps of the high-frequency filtering algorithm based on SCSO optimization described in step (2) are as follows: (21) Take out the load data and record it as x(n); (22) Calculate the frequency axis f of the load data k , that is, where N is the length of the time series data, the formula is as follows: Where: j is the index of the sequence, N is the length of the sequence, t1 and t0 are the first and second time points of the time series data respectively. In order to facilitate the subsequent algorithm optimization, the frequency axis is enlarged 1000 times. (23) Perform fast Fourier transform (FFT) on the load data to obtain the frequency domain signal X k , the formula is as follows: Where: x n The time domain signal representing the load data, X k represents the frequency domain signal of the load data, N represents the length of the signal, and k represents the subscript of the frequency domain signal of the load data. (24) Initialize the filter threshold T, and then calculate the R between the predicted value and the original load 2 The formula is as follows, and this R 2 The value is used as the fitness function of the SCSO algorithm to participate in the optimization and obtain the optimal threshold T best . Where: x represents the original load value, represents the actual load average value, Indicates the load value after filtering at each moment. (25) All frequencies with absolute values ​​greater than the optimal threshold T best The load data is set to zero, and the formula is as follows: X k =0,|f k |>T best (26) Perform inverse Fourier transform (IFFT) on the filtered frequency domain load data to obtain the filtered time domain load data The formula is as follows: The hybrid model load forecasting method based on improved feature processing according to claim 1 is characterized in that: The SMG hybrid prediction model construction of step (3) includes the following steps: (31) Using the periodic feature SD obtained by screening cast , temperature T, solar radiation SR and filtered load data Composition of the dataset Then the dataset is normalized to MIN-MAX to eliminate the impact of different dimensions on deep learning training. (32) A GRU network was built and an MHA layer was added to the GRU network to capture the complex relationship between multiple input features and extract the long-term dependencies in the load data. The SCSO algorithm was then used to optimize the learning rate of deep learning. (33) Using the SMG hybrid forecasting model with optimal parameters, the load forecasting result x is obtained pred .

Images