CN117743829A - Short-term power load quantity prediction method based on deep learning - Google Patents

Abstract

The invention discloses a short-term power load quantity prediction method based on deep learning, which comprises the following implementation steps: (1) performing empirical mode decomposition on load data; (2) sub-sequence classification; (3) high frequency component re-decomposition and parameter optimization; (4) modeling; (5) Compared with the traditional short-term power load prediction method, the method can convert the non-stable and nonlinear original time sequence into a plurality of subsequences and reconstruct the subsequences by introducing modal decomposition, thereby solving the problems of large short-term power load fluctuation and difficult accurate prediction; the high-frequency component is subjected to secondary decomposition, so that the obvious characteristic of time sequence data can be effectively extracted, and the subsequence obtained by secondary decomposition and the subsequence without secondary decomposition are used as model input characteristics, so that the prediction accuracy is effectively improved; the problem of selecting key parameters of LSTM network and VMD decomposition is solved by using SSA to optimize, and the problems of time consumption and poor prediction effect of manual parameter adjustment are solved.

Description

A short-term power load forecasting method based on deep learning

Technical field

The present invention relates to the technical field of power load forecasting, and in particular to a power load forecasting method based on deep learning.

Background technique

Power system load forecasting is to explore the impact of historical changes in power load on future load based on historical data such as power load, economy, society, and meteorology, to seek the intrinsic relationship between power load and various related factors, and to predict future power load. Scientific prediction. Inaccurate power load forecasting may cause power system failures and lead to large-scale power outages, affecting normal production and life of society. Therefore, accurate load forecast modeling is of great significance to ensure the safe and stable operation of the power system.

Traditional load forecasting models lack adaptability and prediction capabilities, have poor robustness, and have inaccurate prediction results, and their accuracy cannot meet the requirements of load forecasting.

Contents of the invention

The technical problem to be solved by this invention is to provide a power load prediction method that combines modal decomposition technology and deep learning to achieve high-precision short-term power load prediction by taking into account both model training time and prediction accuracy. Accurate regulation of electric power load provides certain technical reference. In order to achieve the above-mentioned object of the invention, the technical solutions adopted by the present invention are:

An electric power load forecasting method based on deep learning, which includes the following implementation steps:

(1) Empirical mode decomposition of load data; (2) Subsequence classification; (3) High-frequency component re-decomposition and parameter optimization; (4) Model construction; (5) Result prediction, specifically:

Step 1: Use empirical mode decomposition to decompose the original load data to obtain n intrinsic mode functions IMF _i (i=1, 2,...,n) and residual amounts.

The specific method is:

1) Use cubic splines to interpolate the original signal, find its upper and lower envelopes X _max (t) and X _min (t), and calculate the average envelope as

2) Subtract the original sequence signal from the envelope mean m _j-1 (t), X(t)-m _j-1 (t)=h _j (t), and obtain h _j (t). h _j (t) here is the residual signal.

3) In order to obtain a smoother sequence, repeat the above steps and decompose the remaining components until the termination condition is met.

Step 2: Calculate the sample entropy of the subsequence and classify it.

The IMF components obtained after empirical mode decomposition are arranged from high to low in frequency. Use the sample entropy function to calculate the sample entropy of the intrinsic modal function IMF _i (i=1,2,..,n) and the residual quantity residual, and evaluate the modal component complexity; use the K-means clustering algorithm to divide it into three category, forming high, middle and low frequency components.

Step 3: Use the variational modal VMD algorithm to re-decompose the high-frequency components to obtain a series of modal components.

VMD decomposition is essentially a variational problem that can be solved by establishing the following optimization problem:

Among them, α is the penalty factor; λ is the Lagrangian multiplier operator, u _k is the k-th modal function component; w _k is the center frequency of the k-th modal function; δ (t) is the Dirac function, ||·|| represents the norm, f(t) represents the original signal, and * represents the convolution operator.

Step 4: Establish a corresponding LSTM model for each subsequence for training and prediction, in which the Sparrow search optimization algorithm is used to assist in model parameter selection;

The steps are: first, randomly generate a set of hyperparameters for LSTM model training. Then, the fitness of each set of hyperparameters is determined by evaluating the performance of the LSTM model on the validation set. All sparrows in the sparrow search algorithm are sorted according to the fitness evaluation results, and the sparrows with higher fitness are ranked first. Sparrows are selected for mating and mutation operations based on the ranking results to generate new hyperparameter combinations. The operation is repeated until the algorithm converges, that is, the optimal hyperparameter combination is found. The parameters obtained by the Sparrow search optimization algorithm are sent to LSTM to complete model training.

Step 5: Model prediction and evaluation.

The predicted values obtained by each LSTM prediction model are denormalized, and then the aggregation operation is performed to obtain the final actual predicted value. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R ² ) indicators are introduced to test the prediction effect of the model.

The beneficial effects of the present invention are:

By introducing modal decomposition, the present invention can convert non-stationary and non-linear original time series into several sub-sequences and reconstruct them, solving the problem of large short-term power load fluctuations and difficulty in accurate prediction; by performing secondary decomposition of high-frequency components, It can effectively extract the significant features of time series data, and use the subsequences obtained by secondary decomposition and subsequences without secondary decomposition as model input features, which effectively improves the prediction accuracy; SSA is used to optimize the key parameter selection of LSTM network and VMD decomposition. Solve the problems of time-consuming manual parameter adjustment and poor prediction effect.

Description of drawings

Figure 1 is: a flow chart of the method of the present invention;

Figure 2 is a schematic diagram of classification and reconstruction of subsequences obtained by decomposition in step 2 of the method of the present invention;

Figure 3 is a schematic diagram of the subsequence obtained by using VMD to decompose high-frequency components in the method of the present invention;

Figure 4 is a schematic diagram of the results predicted by the method of the present invention and the true values;

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings: the flow chart of the method of the present invention is shown in Figure 1, and the specific implementation process of the method is as follows:

Step 1: Use empirical mode decomposition (EMD, Empirical Model Decomposition) to decompose the original power load data to obtain several intrinsic mode functions IMF and residual amounts. The specific method is:

1) Use cubic splines to interpolate the original signal, find its upper and lower envelopes X _max (t) and X _min (t), and calculate the average envelope as

2) Subtract the original sequence signal from the envelope mean m _j-1 (t), X(t)-m _j-1 (t)=h _j (t), and obtain h _j (t).

3) In order to obtain a smoother sequence, repeat the above steps and decompose the remaining components until the termination condition is met.

Step 2: Arrange the IMF components obtained after empirical mode decomposition from high to low in frequency. Calculate the sample entropy of all modal components, use the K-means clustering algorithm for classification, and reconstruct the modal components of the same class to form three types of reconstructed components of high, medium and low frequency, as shown in Figure 2.

Sample entropy calculation method:

1) If the original signal is a sequence of length N, respectively x(1), x(2), x(3)...x(N), take samples of length m in order, of which the i-th sample can be Expressed as X(i)=[x(i),x(i+1),L x(i+m-1)], i=1～N-m-1.

2) Calculate the distance d[X(i),X(j)] between samples X(i) and X(j), which is defined as the maximum value of the corresponding element difference, that is The symbol || represents absolute value operation;

3) Given a threshold r, count the number of distances smaller than the threshold r and compare it with Nm, denoted as

4) Solve the sample entropy of the sequence:

Then, based on the calculated sequence sample entropy, KMeans is used to divide the given data set into K clusters (K is set to 3 in the present invention), and the center point corresponding to each sample data is given.

Step 3: Use the VMD algorithm to re-decompose the reconstructed high-frequency components, and use the Sparrow search algorithm to assist parameter optimization, such as the number of decompositions and penalty factors. The specific methods are:

VMD decomposition is essentially a variational problem that can be solved by establishing the following optimization problem:

Among them, α is the quadratic penalty factor; λ is the Lagrange multiplier operator, u _k is the k-th modal function component; w _k is the center frequency of the k-th modal function; α _t is the Dirac function.

Solve the above and get the updated formula as

where n represents the number of iterations. in Respectively expressed/> Fourier transform of f(t), u(t) and λ(t). The update process termination condition is:/> Among them, τ is the noise tolerance, and ε>0 is the convergence threshold.

Step 4: Establish a corresponding LSTM model for each subsequence for training and prediction, in which the Sparrow search optimization algorithm is used to assist in model parameter selection:

The steps are: first, randomly generate a set of hyperparameters for LSTM model training. Then, the fitness of each set of hyperparameters is determined by evaluating the performance of the LSTM model on the validation set. All sparrows in the sparrow search algorithm are sorted according to the fitness evaluation results, and the sparrows with higher fitness are ranked first. Sparrows are selected for mating and mutation operations based on the ranking results to generate new hyperparameter combinations. The operation is repeated until the algorithm converges, that is, the optimal hyperparameter combination is found. The parameters obtained by the Sparrow search optimization algorithm are sent to LSTM to complete model training.

Step 5: Model prediction and evaluation.

The predicted values obtained by each LSTM prediction model are denormalized, and then the aggregation operation is performed to obtain the final actual predicted value. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R ² ) indicators are introduced to test the prediction effect of the model.

The root mean square error reflects the discreteness of the predicted values and the size of the model error.

The mean absolute error is the average of the absolute errors, which can better reflect the actual situation of the predicted value error.

The goodness of fit can reflect the degree to which the independent variable explains the dependent variable. The larger the value, the better the goodness of fit.

Figure 4 is a schematic diagram of the results predicted by the method of the present invention and the actual values. The prediction results (part) are shown in Table 1:

Table 1 Prediction results versus true values

time Actual load kWh Forecast load kWh 22:30 2269.729 2209.407 22:45 2203.306 2116.752 23:00 2385.568 2393.735 23:15 1945.762 1975.920 23:30 1753.181 1724.883

Table 2 is a comparison table of evaluation indicators for the predicted results of the method of the present invention and the comparison model. as follows:

Table 2 Evaluation index table

Model Method of the present invention EMD-LSTM VMD-LSTM R ² 0.921 0.697 0.833 RMSE 128.1 236.98 175.79 MAE 83.65 122.93 100.43

It can be seen from Tables 1 and 2 that when test data is used to predict power load, the prediction results of the method of the present invention have a high degree of fit with the real results, and the prediction accuracy is high.

The deep learning algorithm can mine its inherent laws through data analysis, so as to achieve the purpose of building a model for load forecasting. In view of the time series variation characteristics of the load data, the decomposition algorithm is used to decompose it into multiple modal components of different frequencies, which further reduces the difficulty of data fitting, makes the data curve smooth and periodic, and improves the fitting degree of the neural network.

The above are only preferred embodiments of the present invention. It should be pointed out that those of ordinary skill in the technical field can also make several improvements without departing from the principles of the present invention. These improvements can be made without inventive efforts. The following should also be regarded as the protection scope of the present invention.

Claims (7)

1. A short-term power load prediction method based on deep learning, characterized in that it includes the following implementation steps: (1) Empirical mode decomposition of load data; (2) Subsequence classification; (3) High-frequency components Re-decomposition and parameter optimization; (4) model construction; (5) result prediction. 2. The short-term power load prediction method based on deep learning according to claim 1, characterized in that the specific method steps are: Step (1): Use empirical mode decomposition to decompose the original load data to obtain n original load data. Eigenmode function IMF _i (i=1,2,..,n) and residual amount; Step (2): Calculate the sample entropy of the subsequence and classify it; Arrange the IMF components obtained after empirical mode decomposition from high to low frequency; use the sample entropy function to calculate the samples of the intrinsic mode function IMF _i (i=1,2,..,n) and the residual amount residual Entropy, evaluates the complexity of modal components; uses K-means clustering algorithm to divide them into three categories to form high, medium and low frequency components; Step (3): Use the variational modal VMD algorithm to re-decompose the high-frequency components to obtain a series of modal components; Step (4): Establish a corresponding LSTM model for each subsequence for training and prediction, in which the Sparrow search optimization algorithm is used to assist in model parameter selection; Step (5): Model prediction and evaluation: denormalize the prediction values obtained by each LSTM prediction model, and then perform an aggregation operation to obtain the final actual prediction value. Introduce the root mean square error (RMSE), mean absolute error ( MAE) and coefficient of determination (R ² ) indicators to test the prediction effect of the model. 3. The short-term power load prediction method based on deep learning according to claim 2, characterized in that the specific method of step (1) is: 1) Use cubic splines to interpolate the original signal, find its upper and lower envelopes X _max (t) and X _min (t), and calculate the average envelope as 2) Subtract the original sequence signal from the envelope mean m _j-1 (t), X(t)-m _j-1 (t)=h _j (t), and obtain h _j (t). h _j (t) here is the residual signal; 3) In order to obtain a smoother sequence, repeat the above steps and decompose the remaining components until the termination condition is met. 4. The short-term power load prediction method based on deep learning according to claim 2, characterized in that the specific method of step (3) is: Sample entropy calculation method: 1) If the original signal is a sequence of length N, respectively x(1), x(2), x(3)...x(N), take samples of length m in order, of which the i-th sample can be Expressed as X(i)=[x(i),x(i+1),L x(i+m-1)], i=1～N-m-1; 2) Calculate the distance d[X(i),X(j)] between samples X(i) and X(j), which is defined as the maximum value of the corresponding element difference, that is The symbol || represents absolute value operation; 3) Given a threshold r, count the number of distances smaller than the threshold r and compare it with Nm, denoted as 4) Solve the sample entropy of the sequence: Then, based on the calculated sequence sample entropy, KMeans is used to divide the given data set into K clusters (K is set to 3 in the present invention), and the center point corresponding to each sample data is given. 5. The short-term power load prediction method based on deep learning according to claim 2, characterized in that the specific method of step (3) is: VMD decomposition is essentially a variational problem that can be solved by establishing the following optimization problem: Among them, α is the penalty factor; λ is the Lagrangian multiplier operator, u _k is the k-th modal function component; w _k is the center frequency of the k-th modal function; δ (t) is the Dirac function, ||·|| represents the norm, f(t) represents the original signal, and * represents the convolution operator. 6. The short-term power load prediction method based on deep learning according to claim 2, characterized in that the specific method of step (4) is: First, a set of hyperparameters are randomly generated for LSTM model training; then, the fitness of each set of hyperparameters is determined by evaluating the performance of the LSTM model on the validation set; all sparrows in the sparrow search algorithm are sorted according to the fitness evaluation results, Rank sparrows with higher fitness at the front; select sparrows for mating and mutation operations based on the ranking results to generate new hyperparameter combinations; continue to repeat the operation until the algorithm converges, that is, find the optimal hyperparameter combination; optimize the sparrow search The parameters obtained by the algorithm are sent to LSTM to complete model training. 7. The short-term power load prediction method based on deep learning according to claim 2, characterized in that in step (5): the root mean square error reflects the discreteness of the predicted value and the size of the model error: The mean absolute error is the average of the absolute errors, which can better reflect the actual situation of the predicted value error: The goodness of fit can reflect the degree to which the independent variable explains the dependent variable. The larger the value, the better the goodness of fit: