A power load prediction method based on PSVLBA model

Abstract

The application discloses a power load prediction method based on a PSVLBA model, and comprises the following steps: decomposing historical load by using a PSO-optimized VMD method to obtain sub-sequences of the historical load, grouping and reconstructing each sub-sequence based on frequency size, taking the reconstructed 3D data as the input of an LSTM model, and obtaining a prediction result based on historical data through a full connection layer; according to the strong correlation between power load and date, encoding the date to obtain five date characteristics as the input of a BPNN model, and obtaining a load prediction result based on the date; according to the same period similarity of the power load itself, constructing the power load of the previous week as the same period load as the prediction result, and obtaining the final prediction result based on the Attention mechanism of the three prediction results, so that the PSVLBA model is established, and the future load is accurately predicted. The application can realize fast and accurate prediction of the power load level, and provides a new idea for power load prediction by using a combined model.

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting, and in particular relates to a power load forecasting method based on the PSVLBA model. Background Technology

[0002] The power industry is a crucial foundational sector in the national energy field and a cornerstone for the development of all industries. In the field of power load forecasting, load forecasting based on combined models has yielded a wealth of research results after decades of development.

[0003] In recent years, power load forecasting methods based on integrated models have developed rapidly. Shao B proposed the VMD-WSLSTM network, which uses the VMD method to decompose the original load data and predicts the decomposed components separately. The WSLSTM prediction model built using a weight-sharing mechanism has better training speed and prediction accuracy than LSTM and GRU models. Li S used wavelet transform to decompose the load sequence into a set of different frequency components, and then used an improved ABC algorithm to initialize the weights and biases of the Extreme Learning Machine (ELM). The proposed model has good convergence performance and prediction accuracy. Zheng H proposed the SD-EMD-LSTM model, which decomposes the load sequence through SD feature selection and EMD method, and then builds sub-sequence models through multiple LSTMs. This method has good long-term prediction capability for complex nonlinear power loads. Lian H used ensemble empirical mode decomposition to decompose the original load sequence. Ensemble mode decomposition effectively avoids the mode aliasing phenomenon of empirical mode decomposition and improves the stationarity of sub-sequences. Then, GRU is used to build models for each sub-sequence, and finally a peak correction algorithm is added to correct the predicted values ​​to obtain the final prediction results. Li T uses improved adaptive noisy complete set empirical mode decomposition (CEEMDAN) load data, and then uses the Grey Wolf Optimization (GWO) algorithm on a multi-core extreme learning machine to predict the load.

[0004] For example, Chinese patent document CN110796293A discloses a power load prediction method, which first uses a trained least squares support vector machine model for vertical prediction, then uses a trained capacitive Kalman filter prediction model for horizontal prediction, and finally uses a trained gray neural network model to fuse the two algorithms.

[0005] Chinese patent document CN110991722A discloses a method for predicting power load, including the following steps: S1, constructing a training dataset and a test dataset for the prediction model; S2, establishing a prediction model based on a hierarchical parallel Bayesian neural network; S3, inputting training samples from the training dataset into the prediction model based on the hierarchical parallel Bayesian neural network for training; S4, inputting test data into the trained prediction model for prediction to obtain the power load prediction result.

[0006] However, current power load forecasting methods all suffer from complex model structures and low forecasting accuracy. Summary of the Invention

[0007] This invention provides a power load forecasting method based on the PSVLBA model, which has high forecasting accuracy and a moderate model size.

[0008] A power load forecasting method based on the PSVLBA model includes the following steps:

[0009] (1) Collect power data datasets and preprocess the datasets, including abnormal data detection and correction, and feature normalization;

[0010] (2) Construct the sample set {(X) using the sliding window method t ,Y t )};

[0011] in, Y = (x t ); To process the historical load results of the n1-day prior to the forecast date using the PSO-VMD method; The prediction results of historical loads for n2 days are processed using the SARIMA method; To predict the time characteristics of a day, including year, month, week, day, and holidays; To predict the meteorological characteristics of the day, including temperature, relative humidity, and precipitation; x t To predict the actual daily load;

[0012] (3) The sample set Reconstructed as a trend item Periodic terms and residuals in, After reconstruction in K is the number of decompositions;

[0013] (4) Divide the sample set into training set, validation set and test set according to time sequence and proportion;

[0014] (5) Construct and initialize the PSVLBA model, which includes LSTM, BPNN and Attention connected by fully connected layers;

[0015] (6) Input the training set into the PSVLBA model for training, and input the training set into the model for each sample. Input LSTM, Input BPNN, The input Attention layer produces three sequences of length 24. These three sequences are then concatenated to form a sequence of length 72. Finally, the result of the forward propagation is output through a 72*24 fully connected layer.

[0016] The PSVLBA model was trained using mean squared error (MSE) as the loss function, and the EarlyStopping method was used to optimize the training process.

[0017] (7) Use the trained PSVLBA model for prediction, and input the data to be predicted X. t The predicted daily electricity load is obtained after inputting the data into the model.

[0018] This invention obtains subsequences by performing PSO-optimized VMD decomposition on historical loads, and reconstructs them into high, medium, and low frequency sequences based on frequency magnitude as input to an LSTM model to obtain a prediction output based on historical loads. Since power loads are greatly affected by date features, date features are encoded and used as input to a BPNN model to obtain a prediction output based on date. Because power loads have similar characteristics for the same period, data from the past 5 weeks of the prediction date are constructed and predicted using a SARIMA model. The Attention layer weights the above three prediction results to obtain the final prediction result.

[0019] This invention addresses the task of electricity load forecasting by considering multiple influencing factors and designing a combined model. Compared to using a single deep learning method for forecasting, this model significantly improves forecast accuracy. This invention greatly enhances the accuracy of electricity load forecasting.

[0020] In step (1), the abnormal data detection process is as follows:

[0021] Calculate the average electricity load from Monday to Sunday. and standard deviation σ ij ,

[0022]

[0023]

[0024] In the formula, i represents the day of the week, j represents the time, and k represents the week number. For dates outside the week... Data within the specified range is considered abnormal.

[0025] The formula for feature normalization is as follows:

[0026]

[0027] In the formula, Let j be the normalized value of feature j at time i. Let μ be the value of feature j at time i. j Let σ be the mean of feature j. j Let be the variance of feature j.

[0028] In step (2), the process of using the PSO-VMD method to process the historical load of day n1 before the forecast date is as follows:

[0029] a-1) Initialization parameters: Maximum number of iterations G in the particle swarm optimization algorithm max Particle swarm size m, inertia factor ω; learning factors c1, c2; particle position constraint X max X min Maximum particle velocity V max Variational mode decomposition convergence threshold tol; dual rising parameter tau; frequency initialization method init;

[0030] a-2) Population initialization: Randomly generate m sets of [K, α] as the positions of the particles, and randomly initialize the velocities of the particles;

[0031] a-3) Update the local minimum: Use m sets [K, α] as parameters and... The VMD algorithm is executed as input, calculating the average envelope entropy of all IMFs in each group, and updating the local minimum for each group. The envelope entropy formula is as follows:

[0032]

[0033] a-4) Update the global minimum: Compare the local minimum of each group and update the global minimum;

[0034] a-5) Update the particle positions and velocities of the particle swarm. If the constraints are not met, reset them to the corresponding upper and lower bounds and continue with step a-3) until the iteration is complete. Output the best fitness function value and the corresponding best particle position [K, α].

[0035] In step (2), The construction process is as follows:

[0036] b-1) Definition

[0037] b-2) Long-term trend indicator of electricity load: x0(t)

[0038] x0(t) = D0(t) + x1(t)

[0039] In the formula, D0(t) is the annual index, and x1(t) represents the day of the year;

[0040] b-3) Electricity load cycle variation indicators x1(t), x2(t), x3(t):

[0041] x i (t)=D i (t) / S i i∈{1,2,3}

[0042] S1 is the number of days in the year, S2 is the number of days in the month, S3 is the number of days in the week, and D... i (t)∈{0,1,...,S i -1};

[0043] b-4) Electricity load during holidays: x4(t)

[0044]

[0045] D ± (t)=w + D * (t+1)+w - D * (t-1)+w ± D * (t-1)D * (t+1)

[0046] x4(t)=max(D ± (t),D * (t))

[0047] In the formula w + and w - The influence weights of the day before and the day after the holiday, w ± This reflects the impact of the days before and after holidays.

[0048] In step (4), the training set, validation set, and test set are divided in chronological order at a ratio of 6:2:2.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The algorithm of this invention is simple in principle, easy to implement, and easy to operate.

[0051] 2. This invention reduces the complexity of power load and improves the predictive performance of the model by decomposing historical loads based on the PSO-optimized VMD method.

[0052] 3. This invention encodes date features and establishes a relative relationship between power load and date, thereby improving the predictive performance of the model.

[0053] 4. This invention connects LSTM, BPNN and Attention models through a fully connected layer, combining the advantages of the three models. The corresponding PSVLBA combined model has the highest prediction accuracy. Attached Figure Description

[0054] Figure 1 This is a diagram of Panama's electricity load data in an embodiment of the present invention;

[0055] Figure 2 This is a flowchart illustrating the PSO-VMD method for processing historical loads up to n1 days prior to the forecast date in this embodiment of the invention.

[0056] Figure 3 This is a schematic diagram of the PSVLBA model in this invention;

[0057] Figure 4 This is a comparison chart of the predicted values ​​of the model of this invention and the comparative model;

[0058] Figure 5 This is a regression comparison chart of the predicted values ​​of the model of this invention and the comparative model;

[0059] Figure 6 This is a graph showing the average absolute error of the model of this invention and the comparative model from Monday to Sunday. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0061] The power load forecasting method of the present invention will be described in detail below using the Panama dataset as an example. The power load data for the Panama region in this embodiment is as follows: Figure 1 As shown, this dataset contains electricity load data from 2015 to 2019. The standardized sample length is 35,000, the sampling interval is 1 hour, and the unit is MWh.

[0062] A power load forecasting method based on the PSVLBA model includes the following steps:

[0063] S01, Collect power data dataset, and preprocess the dataset. The preprocessing includes anomaly detection and correction, feature normalization, and selection of data range.

[0064] (1-1) Input power load data and other characteristic data.

[0065] (1-2) Calculate the average electricity load from Monday to Sunday. and standard deviation σ ij

[0066]

[0067]

[0068] In the formula, i represents the day of the week, j represents the time, and k represents the week number. For dates outside the week... Data within the specified range was identified as anomalous. The results of anomalous data processing are shown in Table 1 below.

[0069] Table 1

[0070]

[0071] (1-3) Normalize each feature using a standardization method.

[0072]

[0073] In the formula, Let j be the normalized value of feature j at time i. Let μ be the value of feature j at time i. j Let σ be the mean of feature j. j Let be the variance of feature j.

[0074] S02, using the sliding window method to construct the sample set {(X) t ,Y t )};

[0075] in, Y = (x t ); To process the historical load results of the n1-day prior to the forecast date using the PSO-VMD method; The prediction results of historical loads for n2 days are processed using the SARIMA method; To predict the time characteristics of a day, including year, month, week, day, and holidays; To predict the meteorological characteristics of the day, including temperature, relative humidity, and precipitation; x t To predict the actual daily load.

[0076] like Figure 2 As shown, the specific process of using the PSO-VMD method is as follows:

[0077] a-1) Initialize parameters. Maximum number of iterations G in the particle swarm optimization algorithm. max Particle swarm size m, inertia factor ω; learning factors c1, c2; particle position constraint X max Xmin Maximum particle velocity V max ; Variational mode decomposition convergence threshold tol; Dual rising parameter tau; Frequency initialization method init.

[0078] a-2) Population initialization. Randomly generate m sets of [K, α] as the positions of the particles, and randomly initialize the velocities of the particles.

[0079] a-3) Update local minima. Perform the VMD algorithm on m groups [K, α] to calculate the average envelope entropy of all IMFs in each group, and update the local minima for each group. The envelope entropy formula is as follows:

[0080]

[0081] a-4) Update the global minimum. Compare the local minimums of each group and update the global minimum.

[0082] a-5) Update the particle positions and velocities of the particle swarm. If the constraints are not met, reset them to the corresponding upper and lower bounds and continue with step a-3) until the iteration is complete. Output the best fitness function value and the corresponding best particle position [K, α].

[0083] The construction process is as follows:

[0084] b-1) Definition

[0085] b-2) Long-term trend indicator of electricity load: x0(t)

[0086] x0(t) = D0(t) + x1(t)

[0087] In the formula, D0(t) is the annual index, and x1(t) represents the day of the year;

[0088] b-3) Electricity load cycle variation indicators x1(t), x2(t), x3(t):

[0089] x i (t)…D i (t) / S i i∈{1,2,3}

[0090] S1 is the number of days in the year, S2 is the number of days in the month, S3 is the number of days in the week, and D... i (t)∈{0,1,...,S i -1};

[0091] b-4) Electricity load during holidays: x4(t)

[0092]

[0093] D ± (t)=w + D * (t+1)+w - D * (t-1)+w ± D * (t-1)D * (t+1)

[0094] x4(t)=max(D ± (t),D * (t))

[0095] In the formula w + and w - The influence weights of the day before and the day after the holiday, w ± This reflects the impact of the days before and after holidays.

[0096] S03, will include the sample set Reconstructed as a trend item Periodic terms and residuals in, After reconstruction in K is the number of decompositions.

[0097] S04. Divide the sample set into training set, validation set and test set in chronological order at a ratio of 6:2:2.

[0098] S05, construct and initialize the PSVLBA model, which includes an LSTM, BPNN, and Attention layer connected by fully connected layers. The structure of the PSVLBA model is as follows: Figure 3 As shown.

[0099] S06, input the training set into the PSVLBA model for training, and input each sample's... Input LSTM, Input BPNN, The input Attention layer produces three sequences of length 24. These three sequences are then concatenated to form a sequence of length 72. Finally, the result of the forward propagation is output through a 72*24 fully connected layer.

[0100] The PSVLBA model was trained using mean squared error (MSE) as the loss function, and the EarlyStopping method was used to optimize the training process.

[0101] The prediction errors of the PSVLBA model of this invention and the comparison model are shown in Table 2 below. Figure 4This is a comparison chart of the predicted values ​​of the model of this invention and the comparative model. Figure 5 This is a regression comparison chart of the predicted values ​​of the model of this invention and the comparative model. Figure 6 The graph shows the average absolute error of the model of this invention from Monday to Sunday compared to the comparative model. It can be seen that the model is an improvement over the comparative model, verifying the efficiency and accuracy of this invention in power load forecasting.

[0102] Table 2

[0103]

[0104] Using the method of this invention, fast and accurate load forecasting is achieved for power load forecasting tasks, providing a new approach for power load forecasting using combined models, and has certain reference significance and application value.

[0105] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power load forecasting method based on the PSVLBA model, characterized in that, Includes the following steps: (1) Collect power data datasets and preprocess the datasets, including abnormal data detection and correction, and feature normalization; (2) Construct the sample set {(X) using the sliding window method t ,Y t )}; in, Y = (x t ); To process the historical load results of the n1-day prior to the forecast date using the PSO-VMD method; The prediction results of historical loads for n2 days are processed using the SARIMA method; To predict the time characteristics of a day, including year, month, week, day, and holidays; To predict the meteorological characteristics of the day, including temperature, relative humidity, and precipitation; x t To predict the actual daily load; (3) The sample set Reconstructed as a trend item Periodic terms and residuals in, After reconstruction in K is the number of decompositions; (4) Divide the sample set into training set, validation set and test set according to time sequence and proportion; (5) Construct and initialize the PSVLBA model, which includes LSTM, BPNN and Attention connected by fully connected layers; (6) Input the training set into the PSVLBA model for training, and input the training set into the model for each sample. Input LSTM, Input BPNN, The input Attention layer produces three sequences of length 24. These three sequences are then concatenated to form a sequence of length 72. Finally, the result of the forward propagation is output through a 72*24 fully connected layer. The PSVLBA model was trained using mean squared error (MSE) as the loss function, and the EarlyStopping method was used to optimize the training process. (7) Use the trained PSVLBA model for prediction, and input the data to be predicted X. t The predicted daily electricity load is obtained after inputting the data into the model.

2. The power load forecasting method based on the PSVLBA model according to claim 1, characterized in that, In step (1), the abnormal data detection process is as follows: Calculate the average electricity load from Monday to Sunday. and standard deviation σ ij , In the formula, i represents the day of the week, j represents the time, and k represents the week number. For dates outside the week... Data within the specified range is considered abnormal.

3. The power load forecasting method based on the PSVLBA model according to claim 1, characterized in that, In step (1), the formula for feature normalization is as follows: In the formula, Let j be the normalized value of feature j at time i. Let μ be the value of feature j at time i. j Let σ be the mean of feature j. j Let be the variance of feature j.

4. The power load forecasting method based on the PSVLBA model according to claim 1, characterized in that, In step (2), the process of using the PSO-VMD method to process the historical load of day n1 before the forecast date is as follows: a-1) Initialization parameters: Maximum number of iterations G in the particle swarm optimization algorithm max Particle swarm size m, inertia factor ω; learning factors c1, c2; particle position constraint X max X min Maximum particle velocity V max Variational mode decomposition convergence threshold tol; dual rising parameter tau; frequency initialization method init; a-2) Population initialization: Randomly generate m sets of [K, α] as the positions of the particles, and randomly initialize the velocities of the particles; a-3) Update the local minimum: Use m sets [K, α] as parameters and... The VMD algorithm is executed as input, calculating the average envelope entropy of all IMFs in each group, and updating the local minimum for each group. The envelope entropy formula is as follows: a-4) Update the global minimum: Compare the local minimum of each group and update the global minimum; a-5) Update the particle positions and velocities of the particle swarm. If the constraints are not met, reset them to the corresponding upper and lower bounds and continue with step a-3) until the iteration is complete. Output the best fitness function value and the corresponding best particle position [K, α].

5. The power load forecasting method based on the PSVLBA model according to claim 1, characterized in that, In step (2), The construction process is as follows: b-1) Definition b-2) Long-term trend indicator of electricity load: x0(t) x0(t) = D0(t) + x1(t) In the formula, D0(t) is the annual index, and x1(t) represents the day of the year; b-3) Electricity load cycle variation indicators x1(t), x2(t), x3(t): x i (t)=D i (t) / S i ,i∈{1,2,3} S1 is the number of days in the year, S2 is the number of days in the month, S3 is the number of days in the week, and D... i (t)∈{0,1,...,S i -1}; b-4) Electricity load during holidays: x4(t) D ± (t)=w + D * (t+1)+w _ D * (t-1)+w ± D * (t-1)D * (t-1) x4(t)=max(D ± (t),D * (t)) In the formula w + and w - The influence weights of the day before and the day after the holiday, w ± This reflects the impact of the days before and after holidays.

6. The power load forecasting method based on the PSVLBA model according to claim 1, characterized in that, In step (4), the training set, validation set, and test set are divided in chronological order at a ratio of 6:2:2.

Images