Short-term load prediction method and system based on data decomposition and deep learning

Abstract

The invention discloses a short-term load prediction method and system based on data decomposition and deep learning, and the method comprises the following steps: S1, obtaining historical load data and feature data, detecting an abnormal value through employing a box plot theory, and carrying out the mean value filling of the abnormal value through employing the data at the same time of a previous day and a next day; s2, reducing the unpredictability of the load data through variational mode decomposition, and obtaining intrinsic mode components of different frequencies; s3, analyzing the correlation between the feature data and the load based on the maximum information coefficient, realizing feature dimension reduction, and constructing an input data set; s4, constructing a GRU prediction model, carrying out normalization processing on an input data set, and dividing the input data set into a training set and a test set according to a fixed proportion; and S5, training the GRU model to realize data prediction, and carrying out reverse normalization on a prediction result. According to the method, nonlinearity and non-stationarity of power load data are considered, and the short-term load prediction precision can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power load demand prediction, in particular to a short-term load prediction method and system based on data decomposition and deep learning. BACKGROUND

[0002] With the development of new power systems and the widespread application of renewable energy, load prediction has become a challenge. The instability and seasonality of renewable energy sources such as wind and solar energy make it essential to predict load changes in advance.

[0003] Through accurate load prediction, the power system can be better planned and managed, reducing the risk of overload and energy waste, and improving the stability and reliability of the power grid.

[0004] In load prediction, accuracy is the key to effective prediction. Patent 202311480541.4 proposes using transformer for power load demand prediction, which can achieve high-precision prediction, but transformer model is mainly applied to long-time step prediction, and the prediction effect in short-term load prediction needs to be further improved. Patent 202311496873.1 uses LSTM as a prediction model to capture high-correlation feature information, improving the accuracy and effectiveness of model prediction, but the LSTM model structure is complex and involves many parameters.

[0005] In summary, the present application combines data decomposition technology and deep learning model and proposes a load prediction method and system that enhances prediction accuracy and can adapt to different time scales of load prediction. SUMMARY

[0006] The present application addresses the shortcomings of the prior art and provides a short-term load prediction system and method based on data decomposition and deep learning.

[0007] To solve the above technical problems, the present application solves the problems through the following technical solutions:

[0008] The short-term load prediction method and system based on data decomposition and deep learning, the method comprising the following steps:

[0009] S1: Obtain historical load data and feature data, apply box plot theory to detect outliers, and use the same time data of the previous day and the next day to fill in the mean value of the outliers;

[0010] S2: Use variational mode decomposition to reduce the unpredictability of load data, and obtain intrinsic mode components of different frequencies;

[0011] S3: Based on the maximum information coefficient, analyze the correlation between feature data and load, realize feature dimension reduction, and construct an input data set;

[0012] S4: Construct a GRU prediction model, and normalize the input data set, and divide it into training set and test set according to fixed proportion;

[0013] S5: Train the GRU model, then realize data prediction, and reverse normalize the prediction result.

[0014] As preferred, in S1, due to the failure of the data acquisition system and the transmission delay, the original data set is abnormal, the abnormal value is detected by the box plot theory, the mean value is filled in by using the load data of the previous day t time and the load data of the next day t time, and the complete data set is constructed, and the specific calculation is as follows:

[0015]

[0016] Wherein, x(d,t) represents the load data of the dth day t time; x(d+1,t) represents the load data of the d+1th day t time; x(d-1,t) represents the load data of the d-1th day t time.

[0017] As preferred, in S2, the power load data has strong nonlinearity and non-stationary characteristics, and the variational mode decomposition can effectively reduce its unpredictability, and the specific calculation is as follows:

[0018] S21: Construct a constrained optimization model as follows:

[0019]

[0020] Wherein, y(t) is the original sequence, u k (t) is the kth subsequence, ω k is the center frequency of the kth subsequence, K is the mode decomposition number, δ(t) is the Dirac function, * is the convolution operation, is the partial derivative, j is the imaginary unit, t is the time unit.

[0021] S22: Based on the quadratic penalty factor and the Lagrange multiplier, the constrained optimization model is transformed, and the result is as follows:

[0022]

[0023] S23: Iteratively update u k , ω k and λ, λ(t) is the Lagrange multiplier, α is the quadratic penalty factor, ω is the frequency, and termination judgment is performed, and the calculation is as follows:

[0024]

[0025] Wherein, ε is the discrimination parameter, n is the iteration number, is the corresponding Fourier transform, is the corresponding Fourier transform.

[0026] As preferred, in S3, the correlation between the feature data and the load is explored based on the maximum information coefficient, specifically as follows:

[0027] S31: Calculate mutual information:

[0028]

[0029] where I[X, Y] is the mutual information, p[x, y] is the joint probability density function, and p(x) and p(y) are the probability density functions, respectively;

[0030] S32: Calculate the maximum information coefficient:

[0031]

[0032] where MIC[X, Y] is the maximum information coefficient, |X| is the number of grids divided in the X direction, |Y| is the number of grids divided in the Y direction, and B is the grid limit variable;

[0033] In S4, the principle of the GRU model is as follows:

[0034] z t = σ(ω z x t + u z h t-1 + b z )

[0035] r t = σ(ω r x t + u r h t-1 + b r )

[0036]

[0037] where r t is the reset gate, z t is the update gate, h t-1 is the hidden state at t-1, h t is the hidden state at t, h t is the candidate hidden state at t, x z is the input at t, ω r , ω h , and ω z are weights, u r , u h , and u z are weights.is a weight; b z is a weight; b r is a weight; b h is a bias.

[0038] The short-term load prediction system based on data decomposition and deep learning comprises:

[0039] The power load data preprocessing module is used for identifying and filling the outliers in the data set to obtain a complete data set.

[0040] The power load data decomposition module is used for decomposing the load of the complete data set in the power load data preprocessing module to obtain intrinsic mode components of different frequencies.

[0041] The feature engineering module is used for analyzing the correlation between the load of the complete data set in the power load data preprocessing module and the feature data, and then realizing feature dimension reduction.

[0042] The power load prediction module is used for inputting the data obtained by the power load data decomposition module and the feature engineering module into a GRU model for training, and obtaining a final load prediction result.

[0043] The present application has the following technical effects:

[0044] The present application combines data decomposition technology and deep learning model, and proposes a load prediction method and system, which enhances the prediction accuracy and can adapt to load prediction of different time scales. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flowchart of the inventive method. DETAILED DESCRIPTION

[0046] The present application will be further described in detail below in combination with the accompanying drawings and embodiments. Figure 1

[0047] Embodiment 1

[0048] The short-term load prediction method and system based on data decomposition and deep learning comprise the following steps:

[0049] S1: Obtain historical load data and feature data, apply box plot theory to test outliers, and use the same time data of the previous day and the next day to fill the outliers with mean value.

[0050] S2: Use variational mode decomposition to reduce the unpredictability of the load data to obtain intrinsic mode components of different frequencies.

[0051] ​S3: Based on the maximum information coefficient analysis of the correlation between feature data and load, dimensionality reduction is realized, and an input data set is constructed;

[0052] S4: A GRU prediction model is constructed, and the input data set is normalized and divided into training set and test set according to a fixed proportion;

[0053] S5: The GRU model is trained to realize data prediction, and the prediction result is denormalized.

[0054] In S1, due to the fault and transmission delay of the data acquisition system, the original data set is abnormal, the abnormal value is detected by the box plot theory, the mean value is filled in by using the load data of the previous day t time and the load data of the next day t time, and the complete data set is constructed. The specific calculation is as follows:

[0055]

[0056] Among them, x(d,t) represents the load data of the dth day t time; x(d+1,t) represents the load data of the d+1th day t time; x(d-1,t) represents the load data of the d-1th day t time.

[0057] In S2, the power load data has strong nonlinearity and non-stationary characteristics, and the variational mode decomposition can effectively reduce its unpredictability, as follows:

[0058] S21: The constrained optimization model is constructed as follows:

[0059]

[0060] Among them, y(t) is the original sequence, u k (t) is the kth subsequence, ω k is the center frequency of the kth subsequence, K is the mode decomposition number, δ(t) is the Dirac function, * is the convolution operation, is the partial derivative, j is the imaginary unit, t is the time unit;

[0061] S22: Based on the quadratic penalty factor and the Lagrange multiplier, the constrained optimization model is transformed, and the result is as follows:

[0062]

[0063] S23: Update u k , ω k and λ, λ(t) is the Lagrange multiplier, α is the quadratic penalty factor, ω is the frequency, and the termination criterion is performed, and the calculation is as follows:

[0064]

[0065] Where ε is the discriminant parameter and n is the number of iterations. for The corresponding Fourier transform, for The corresponding Fourier transform.

[0066] In S3, the correlation between feature data and load is explored based on the maximum information coefficient, as follows:

[0067] S31: Calculate mutual information:

[0068]

[0069] Where I[X,Y] represents mutual information, p(x,y) is the joint probability density function, and p(x) and p(y) are the marginal probability density functions, respectively;

[0070] S32: Calculate the maximum information coefficient:

[0071]

[0072] Where MIC[X,Y] is the maximum information coefficient, |X| is the number of grids in the X direction, |Y| is the number of grids in the Y direction, and B is the grid constraint variable;

[0073] In S4, the GRU model works as follows:

[0074] z t =σ(ω) z x t +u z h t-1 +b z )

[0075] r t =σ(ω) r x t +u r h t-1 +b r )

[0076]

[0077] Where, r t To reset the door, z t To update the gate, h t-1 Let h be the hidden state at time t-1. t Let the hidden state be at time t. Let x be the candidate hidden state at time t. t Let ω be the input at time t. z ω r and ω h As the weight, u z, u r and u h are weights; b z , b r and b h are biases.

[0078] Select the annual load of a certain area as the verification object, and take RMSE, MAE and MAPE as the measurement indexes, and calculate as follows:

[0079]

[0080] In the formula, N is the sample capacity; y t ′ is the actual value; y t ″ is the predicted value.

[0081] Table 1 different model prediction result table

[0082] Prediction model RMSE MAE MAPE RNN 198.718 156.933 1.898 LSTM 181.439 136.861 1.662 GRU 148.370 115.571 1.422

[0083] As can be seen from Table 1, compared with RNN and LSTM, the RMSE based on GRU is reduced by 25.34% and 18.23%, the MAE is reduced by 26.36% and 15.56%, and the MAPE is reduced by 25.08% and 14.44%.

[0084] Table 2 decomposition method prediction result table

[0085] Decomposition method RMSE MAE MAPE EMD 171.452 130.134 1.648 CEEMDAN 167.339 126.917 1.545 VMD 148.370 115.571 1.422

[0086] As can be seen from Table 2, compared with EMD decomposition and CEEMDAN decomposition, the RMSE based on VMD decomposition is reduced by 13.46% and 11.34%, the MAE is reduced by 11.19% and 8.94%, and the MAPE is reduced by 13.71% and 7.96%.

[0087] Embodiment 2

[0088] The short-term load prediction system based on data decomposition and deep learning comprises:

[0089] The power load data preprocessing module is used for identifying and filling the outliers in the data set to obtain a complete data set.

[0090] The power load data decomposition module is used for decomposing the load of the complete data set in the power load data preprocessing module to obtain intrinsic mode components of different frequencies.

[0091] The feature engineering module is used for analyzing the correlation between the load of the complete data set in the power load data preprocessing module and the feature data, and then realizing feature dimension reduction.

[0092] The power load prediction module, the power load data decomposition module and the feature engineering module are connected with the power load prediction module, and the power load prediction module is used for inputting the data obtained by the power load data decomposition module and the feature engineering module into the GRU model for training, and obtaining a final load prediction result.

Claims

1. A short-term load forecasting method based on data decomposition and deep learning, characterized in that, The method comprises the following steps: S1: Obtain historical load data and feature data, apply box plot theory to detect outliers, and use the same time data of the previous day and the next day to fill in the mean value of the outliers; S2: Use variational mode decomposition to reduce the unpredictability of the load data, and obtain intrinsic mode components of different frequencies; S3: Based on the maximum information coefficient, analyze the correlation between the feature data and the load, realize feature dimension reduction, and construct an input data set; S4: Construct a GRU prediction model, normalize the input data set, and divide it into a training set and a test set according to a fixed proportion; S5: Train the GRU model, then realize data prediction, and reverse normalize the prediction results. 2.The short-term load forecasting method based on data decomposition and deep learning according to claim 1, characterized in that, In S1, due to the failure of the data acquisition system and transmission delay, the original data set is abnormal, the outliers are detected by box plot theory, the mean value is filled in by using the load data at time t of the previous day and the load data at time t of the next day, and a complete data set is constructed. The specific calculation is as follows: Where x(d,t) represents the load data at time t on the dth day; x(d+1,t) represents the load data at time t on the d+1th day; and x(d-1,t) represents the load data at time t on the d-1th day. 3.The short-term load forecasting method based on data decomposition and deep learning according to claim 1, characterized in that, In S2, the power load data has strong nonlinearity and non-stationary characteristics, and variational mode decomposition can effectively reduce its unpredictability, as follows: S21: Construct a constrained optimization model as follows: where y(t) is the original sequence, u k (t) is the kth subsequence, ω k is the center frequency of the kth subsequence, K is the number of modal decomposition, δ(t) is the Dirac function, * is the convolution operation, is the partial derivative, j is the imaginary unit, and t is the time unit. S22: Based on the quadratic penalty factor and the Lagrange multiplier, the constrained optimization model is transformed as follows: S23: iteratively update u k , ω k and λ, λ(t) is a Lagrange multiplier, a is a quadratic penalty factor, ω is a frequency, and a termination determination is made, which is calculated as follows: Wherein, ε is a discrimination parameter, n is the iteration number, is the Fourier transform corresponding to is the Fourier transform corresponding to 4.The short-term load forecasting method based on data decomposition and deep learning according to claim 1, characterized in that, In S3, the correlation between the feature data and the load is explored based on the maximum information coefficient, as follows: S31: Calculate mutual information: Where I[X,Y] is mutual information, p[x,y] is joint probability density function, p(x) and p(y) are probability density functions; S32: Calculate the maximum information coefficient: Where MIC[X,Y] is the maximum information coefficient, |X| is the number of grids divided in the X direction, |Y| is the number of grids divided in the Y direction, and B is the grid limit variable. 5.The short-term load forecasting method based on data decomposition and deep learning according to claim 1, characterized in that, In S4, the principle of the GRU model is as follows: z t = σ(ω z x t + u z h t-1 + b z ) r t = σ(ω r x t + u r h t-1 + b r ) where r t is a reset gate, z t is an update gate, h t-1 is the hidden state at time t-1, h t is the hidden state at time t, is a candidate hidden state at time t, x t is the input at time t, ω z , ω r and ω h are weights, u z , u r and u h are weights; b z , b r and b h are biases.

6. A short-term load forecasting system based on data decomposition and deep learning, characterized by, It includes: A power load data preprocessing module for identifying and filling in outliers in the data set to obtain a complete data set; A power load data decomposition module for decomposing the load of the complete data set in the power load data preprocessing module to obtain intrinsic mode components of different frequencies; A feature engineering module for analyzing the correlation between the load and the feature data of the complete data set in the power load data preprocessing module, and then realizing feature dimension reduction; A power load prediction module for inputting the data obtained by the power load data decomposition module and the feature engineering module into the GRU model for training, and obtaining the final load prediction result.

Images