Short-term power load prediction method fusing MIC feature optimization and BWO-CNN-BiGRU

Abstract

The invention discloses a short-term power load prediction method fusing MIC feature optimization and BWO-CNN-BiGRU, and the method comprises the following steps: S1, obtaining power load data of a certain region, including power load data and meteorological feature data, screening meteorological feature data with a high degree of association with a power load through MIC, and carrying out the selection of the meteorological feature data through the MIC; selecting the first five kinds of feature data with relatively high MIC values and the power load data to form a fusion data set; s2, performing data preprocessing on the fused data set; s3, a BWO-CNN-BiGRU load prediction model is constructed; and S4, performing short-term power load prediction by using the trained prediction model, comparing with other prediction models, and evaluating the prediction effect. The short-term power load prediction method combines the maximum information coefficient and the deep learning model to predict the short-term power load, and has the advantages of good prediction precision and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU. Background Art

[0002] With the development of various energy-saving technologies, accurate load forecasting plays an increasingly important role in energy conservation management. In recent years, load forecasting technology has received increasing attention. Usually, load forecasting includes long-term load forecasting (LTLF) for loads over one year, medium-term load forecasting (MTLF) for loads from a few weeks to a year, short-term load forecasting (STLF) for loads from one day to a week, and very short-time load forecasting (VSTLF) for loads from a few minutes to a few hours. LTLF and MTLF can estimate the trend of load changes and are suitable for long-term planning of the system during the design stage. STLF and VSTLF can generate precise control and scheduling load requirements and are more suitable for short-term control of existing systems.

[0003] Traditional load forecasting methods are mainly based on statistical models to learn the nonlinear relationship between various influencing factors and loads, but the fitting effect is not ideal. With the continuous development of artificial intelligence technology, the use of neural networks to establish nonlinear black box fitting models between input and output has shown obvious advantages. The application of neural networks for short-term load forecasting can significantly improve the prediction accuracy. The power sector can rely on accurate load forecasting to reasonably arrange power transmission and distribution, accurately control operating costs, and improve the economic benefits of unit operation while ensuring the stable operation of the power grid. Therefore, based on previous research, it is of great practical significance to further explore a high-precision short-term load forecasting model. Summary of the invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU, which has the advantages of good diagnostic accuracy and robustness.

[0005] The present invention proposes a short-term power load forecasting method that integrates MIC feature optimization and BWO-CNN-BiGRU, and the method steps are as follows:

[0006] S1: Obtain power load data, use MIC to filter meteorological characteristic data associated with the power load data, and form a fusion data set with the filtered meteorological characteristic data and power load data;

[0007] S2: perform data preprocessing on the fused data set;

[0008] S3: Build a BWO-CNN-BiGRU load forecasting model to perform short-term power load forecasting.

[0009] Preferably, the method steps of using MIC to screen meteorological characteristic data associated with power load data in S1 are as follows:

[0010] S11: Calculate the mutual information MI between L and C based on the predicted electric load variable L and weather variable C

[0011]

[0012] Where p(l,c) is the joint probability density of L and C, p(l) and p(c) are the marginal probability densities of L and C respectively;

[0013] S12: Given a pair of integers (x, y), define a grid G(x, y), calculate the mutual information between L and C in the current grid G, and traverse all grids to obtain the maximum MI value for (x, y)

[0014] f' MI (D,G(x,y))=maxf MI (D,G(x,y))

[0015] In the formula, f MI (D, G(x, y)) is the MI value of the data set D = (L, C) on G(x, y);

[0016] S13: Normalize the maximum MI value in different grids. The normalized value is between (0,1). The normalization formula is:

[0017]

[0018] S14: Take all M G The maximum value of (xy) is taken as the MIC value of the load variable and the weather variable.

[0019]

[0020] Where B(n) is a function of n.

[0021] Preferably, the data preprocessing method for the fused data set in S2 is: filling missing values ​​in the data set with a linear interpolation method and normalizing the data.

[0022] Preferably, the method steps for constructing the BWO-CNN-BiGRU load forecasting model in S3 are as follows:

[0023] S31: Use a sliding window to dynamically take values ​​of the original data, and then construct an input feature map of the CNN feature extraction layer. Use the convolutional neural network CNN to extract a multi-dimensional feature vector that represents the load change, and construct a time series as the sample data of the input layer of the bidirectional gated recurrent unit network prediction model BiGRU;

[0024] S32: Use the BiGRU model for bidirectional loop training to learn the positive and negative laws of load data. Each GRU layer maintains its own hidden state and updates it according to the current input and the hidden state of the previous time step;

[0025] S33: The Beluga optimization algorithm (BWO) is used to optimize the model’s weights and other hyperparameters to further improve the prediction model performance, and finally a BWO-CNN-BiGRU load prediction model is constructed.

[0026] Preferably, the internal hidden layer unit calculation formula of the BiGRU model in S32 is as follows:

[0027]

[0028] In the formula, x t represents the state at time t; and Represents the forward and backward states of the hidden layer at time t; and Represents the forward and backward states of the hidden layer at time t; and represents the forward and backward weights of the hidden layer at time t; b t is the bias of the hidden layer state at time t.

[0029] Preferably, the method steps of optimizing selection by the White Whale Optimization Algorithm BWO in S33 are as follows:

[0030] S331: Initialization is performed, and the population initialization position is:

[0031]

[0032] Where n represents the size of the beluga population; d represents the dimension of the design variable;

[0033] The fitness values ​​for all beluga whales are:

[0034]

[0035] S332: Entering the exploration phase, the mathematical model established by the swimming behavior of the beluga whale, that is, the beluga whale position update formula is as follows:

[0036]

[0037] In the formula, T represents the current number of iterations; represents the j-dimensional updated position of the i-th beluga whale, p j is a random number randomly selected from the d-dimensional space, (j=1,2,..d); Indicates that the i-th beluga whale is at p j Position on the dimension; and Represent the current positions of the i-th and r-th beluga whales, respectively. r represents a randomly selected beluga whale. r1 and r2 represent random numbers between (0,1). sin(2πr2) and cos(2πr2) indicate that the mirrored beluga whale fins are facing the water surface.

[0038] S333: Entering the development stage, the Levy flight strategy was introduced to capture prey. The mathematical model of predation is expressed as:

[0039]

[0040] In the formula, T represents the current number of iterations; and They represent the i-th beluga whale at the current position and a randomly selected beluga whale respectively; Indicates the updated position of the i-th beluga whale; represents the optimal position of the white whale; r3 and r4 represent random numbers between (0,1) respectively; C1 represents the random jump intensity of Levy flight, C1=2r4(1-T / T max );L F is Levy's flight function;

[0041] S334: Entering the whale fall phase, the whale position is updated by using the white whale position and the whale fall step length. The formula is as follows:

[0042]

[0043] Where r5, r6 and r7 are random numbers between (0,1); X step represents the step length of the whale's fall, namely:

[0044] X step =(u b -l b )exp(-C2T / T max )

[0045] In the formula, T and T max Respectively represent the current number of iterations and the maximum number of iterations; u b and l b Represent the upper and lower bounds of the search space respectively; C2 represents the step size factor related to the probability of whale falling and the size of the population (C2 = 2Wf ×n); the probability of a whale falling is W f is a linear function, that is:

[0046] W f =0.1-0.5T / T Max

[0047] where the probability of a whale falling decreases from 0.1 in the initial iteration to 0.05 in the last iteration.

[0048] Preferably, L F The calculation method is as follows:

[0049]

[0050] Where u and v are normally distributed random numbers; β is a constant, β = 1.5.

[0051] Beneficial technical effects of the present invention:

[0052] (1) To solve the problem that the prediction accuracy is affected by data redundancy due to multiple feature data, the present invention first performs MIC feature optimization on the feature data such as meteorological data, and then forms a fused data set with the power load data.

[0053] (2) In view of the characteristics of the power load time series, the present invention introduces a CNN neural network model to extract the nonlinear characteristics of the sequence, and inputs the features into the input layer of the bidirectional gated recurrent unit network prediction model BiGRU as features. The BiGRU model is used for bidirectional cyclic training to learn the positive and reverse rules of the load data, which can dig deeper into the historical characteristics of the load time series and improve the prediction accuracy.

[0054] (3) Aiming at the characteristics of neural network models that are prone to local optimization and difficulty in finding parameters, the present invention introduces the Beluga optimization algorithm BWO to optimize the parameters of CNN-BiGRU and establishes a load prediction model based on BWO-CNN-BiGRU. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is the overall framework diagram of the short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU proposed in the present invention;

[0056] Figure 2 This is a bar chart showing the importance ranking between feature factors and loads in the short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU proposed in the present invention;

[0057] Figure 3It is an optimization flow chart of the BWO algorithm in the short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU proposed in the present invention;

[0058] Figure 4 This is a comparison chart of the short-term power load forecasting method of integrating MIC features with BWO-CNN-BiGRU proposed in the present invention and other forecasting models. DETAILED DESCRIPTION

[0059] The present invention will be further explained below in conjunction with specific embodiments.

[0060] Reference Figure 1 The present invention proposes a short-term power load forecasting method that integrates MIC feature optimization and BWO-CNN-BiGRU. The method steps are as follows:

[0061] S1: Obtain the power load data of a certain area, use MIC to screen the meteorological characteristic data with a high degree of correlation with the power load, and select the top five characteristic data with higher MIC values ​​and the power load data to form a fusion data set;

[0062] Among them: the power load data is the power load data and meteorological characteristic data (temperature, humidity, wind speed, perceived temperature, sunshine and rainfall, as well as the calculated daily maximum temperature, minimum temperature, average temperature, average perceived temperature and average rainfall) collected every 15 minutes from 96 points every day in a certain area of ​​Shandong for a total of 375 days.

[0063] The specific steps of using MIC to screen meteorological characteristic data with a high degree of correlation with power load are as follows:

[0064] S11: Assume the predicted electric load variable L and the weather variable C. Calculate the mutual information MI between L and C, and the calculation formula is:

[0065]

[0066] Where p(l,c) is the joint probability density of L and C, p(l) and p(c) are the marginal probability densities of L and C respectively.

[0067] S12: Given a pair of integers (x, y), define a grid G(x, y), calculate the mutual information between L and C in the current grid G, and traverse all grids to obtain the maximum MI value for (x, y):

[0068] f' MI (D,G(x,y))=maxf MI (D,G(x,y))

[0069] In the formula, f MI(D, G(x, y)) is the MI value of the data set D = (L, C) on G(xy).

[0070] S13: Normalize the maximum MI value in different grids. The normalized value is between (0,1). The normalization formula is:

[0071]

[0072] S44: Take all M G(x,y) The maximum value is taken as the MIC value of the load variable and the weather variable. The calculation formula is:

[0073]

[0074] In the formula, B(n) is a function about n, set n 0.6 .

[0075] S2: Data preprocessing for fusion dataset

[0076] The data preprocessing method for the fused data set is: fill the missing values ​​in the data set with linear interpolation, and normalize the data using Z-score normalization. The specific calculation method and formula are as follows:

[0077]

[0078] Where: x is the original value; x′ is the standardized value; m(x) is the average value of all data corresponding to the feature x; s td is the standard deviation of all data corresponding to the feature x.

[0079] S3: Construct the BWO-CNN-BiGRU load forecasting model. The specific construction process is as follows:

[0080] S31: In order to increase the nonlinear feature extraction process, the sliding window is used to dynamically take values ​​of the original data, and then the input feature map of the CNN feature extraction layer is constructed. The multidimensional feature vector representing the load change is extracted using the convolutional neural network CNN, and the time series is constructed as the sample data of the input layer of the bidirectional gated recurrent unit network prediction model BiGRU; specifically, the sliding window is used to slide and take values ​​on the time dimension of the original data set. The dimension size of the original data set is m×n (n features, and the length of each feature sequence is m). The i-th feature sequence can be expressed as:

[0081] X i (m) = {x i (1),x i (2),…,x i (m)}

[0082] In the formula, x i (k) represents the observed value of the i-th feature at the k-th moment;

[0083] If the kth moment is taken as the current moment, the tth moment before the i-th feature can be recorded as:

[0084] X i ′={x i (k-t+1),…,x i (k-1),x i (k)}

[0085] In the formula, t is the sliding window size; then the window is moved according to the sliding stride size to dynamically take values ​​of the original data set, and the data set after taking values ​​through the sliding window is used as the input feature map of the CNN model.

[0086] S32: Use the BiGRU model for bidirectional loop training to learn the positive and negative laws of load data. Each GRU layer maintains its own hidden state and updates it according to the current input and the hidden state of the previous time step. The internal hidden layer unit calculation formula is as follows:

[0087]

[0088] Where: x t represents the state at time t; and Represents the forward and backward states of the hidden layer at time t; and Represents the forward and backward states of the hidden layer at time t; and represents the forward and backward weights of the hidden layer at time t; b t is the bias of the hidden layer state at time t.

[0089] S33: The weights and other hyperparameters of the model are optimized and selected through the Beluga optimization algorithm BWO to further improve the performance of the prediction model, and finally the BWO-CNN-BiGRU load prediction model is constructed. The Beluga optimization algorithm BWO is a population-based mechanism, which assumes each beluga as a candidate solution and continuously updates it during the optimization process. First, it is initialized, and the population initialization position is:

[0090]

[0091] Where n represents the size of the beluga whale population; d represents the dimension of the design variable.

[0092] The fitness values ​​for all beluga whales are as follows:

[0093]

[0094] The BWO algorithm can switch from exploration to development, depending on the balance factor B f ,Right now:

[0095] B f =B0(1-T / 2T max )

[0096] Where, T represents the current number of iterations; T max represents the maximum number of iterations; B0 represents a random number between (0,1) in each iteration; the exploration phase occurs at the balance factor B f >0.5, while the development phase occurs in B f ≤0.5. As the number of iterations T increases, B f The fluctuation range of decreases from (0,1) to (0,0.5), so the occurrence probabilities of the development and exploration stages change significantly, and the probability of the development stage increases with the increasing number of iterations T.

[0097] Entering the exploration stage, the mathematical model established by the swimming behavior of the beluga whale, that is, the beluga whale position update formula is as follows:

[0098]

[0099] In the formula, T represents the current number of iterations; represents the j-dimensional updated position of the i-th beluga whale, p j is a random number randomly selected from the d-dimensional space, (j=1,2,..d); Indicates that the i-th beluga whale is at p j Position on the dimension; and denote the current positions of the i-th and r-th beluga whales respectively, r denotes a randomly selected beluga whale; r1 and r2 denote random numbers between (0,1); sin(2πr2) and cos(2πr2) denote that the mirrored beluga fins are facing the water surface; the dimensions are selected according to odd and even numbers, and the updated positions reflect the synchronization or mirroring behavior of the beluga whales when swimming or diving; the two random numbers r1 and r2 are used to enhance the random operator in the exploration phase.

[0100] Entering the development stage, the Levy flight strategy was introduced to capture prey, and its mathematical model of predation is expressed as:

[0101]

[0102] In the formula, T represents the current number of iterations; and They represent the i-th beluga whale at the current position and a randomly selected beluga whale respectively; Indicates the updated position of the i-th beluga whale; represents the optimal position of the white whale; r3 and r4 represent random numbers between (0,1) respectively; C1 represents the random jump intensity of Levy flight, C1=2r4(1-T / T max ). F is Levy's flight function, which is calculated as follows:

[0103]

[0104] Where u and v are normally distributed random numbers; β is a constant, β = 1.5.

[0105] Finally, we enter the whale fall stage and update the whale position by using the white whale position and the whale fall step length. The formula is as follows:

[0106]

[0107] Where r5, r6 and r7 are random numbers between (0,1); X step Represents the step length of the whale's fall, namely:

[0108] X step =(u b -l b )exp(-C2T / T max )

[0109] In the formula, T and T max Respectively represent the current number of iterations and the maximum number of iterations; u b and l b Represent the upper and lower bounds of the search space respectively; C2 represents the step size factor related to the probability of whale falling and the size of the population (C2 = 2W f ×n); the probability of a whale falling is W f is a linear function, that is:

[0110] W f =0.1-0.5T / T Max

[0111] Here, the probability of the whale falling decreases from 0.1 in the initial iteration to 0.05 in the last iteration, indicating that the beluga whale is less dangerous when it gets closer to the food source during the optimization process.

[0112] S4: Use the trained prediction model to perform short-term power load forecasting, compare it with other prediction models, and evaluate the prediction results.

[0113] Among them: the test set is input into the trained BWO-CNN-BiGRU model to predict the power load, and compared with BP, GRU, BiGRU, and CNN-BiGRU. The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to evaluate the prediction effect of the model to verify the effectiveness and superiority of the proposed method. The experimental results are shown in Table 1.

[0114] The root mean square error (RMSE) calculation formula is:

[0115]

[0116] The mean absolute percentage error (MAPE) is calculated as:

[0117]

[0118] Where: n is the number of samples; y rea is the actual value of the sample; pre is the load forecast value of the model

[0119] Table 1 Comparison of evaluation indicators

[0120] Prediction Model RMSE MAPE BP 0.3886 0.8731 GRU 0.1712 0.5096 BiGRU 0.1187 0.3960 CNN-BiGRU 0.1032 0.3585 BWO-CNN-BiGRU 0.0985 0.1909

[0121] It can be seen from the test results in Table 1 that the RMSE and MAPE of the BWO-CNN-BiGRU prediction model proposed in the present invention are both smaller than those of other models, indicating that the prediction error is minimal, the accuracy is highest, and it has better prediction accuracy, which is significantly better than other prediction models.

[0122] Reference Figure 2 , which is a bar chart of the importance ranking between characteristic factors and load in the short-term power load forecasting method of integrating MIC feature optimization and BWO-CNN-BiGRU proposed in the present invention; wherein S1~S10 respectively correspond to temperature, humidity, wind speed, perceived temperature, sunshine and rainfall, and the calculated daily maximum temperature, minimum temperature, average temperature, average perceived temperature and average rainfall, and the top five factors with higher MIC values ​​are selected.

[0123] Reference Figure 3 , which is the optimization flow chart of the BWO algorithm. By simulating the population action of beluga whales, the population is initialized first, and then the fitness and balance factor B are updated. f , the exploration phase occurs at the equilibrium factor B f >0.5, while the development phase occurs in B f ≤0.5. Eventually, it enters the whale fall phase and repeats the process until the maximum number of iterations is reached.

[0124] Reference Figure 4, which is a comparison chart of the short-term power load forecasting method of integrating MIC features with BWO-CNN-BiGRU proposed in the present invention and other forecasting models. It can be seen that the model of the present invention is more accurate than other models.

Claims

1. A short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU, characterized in that: The steps are as follows: S1: Obtain power load data, use MIC to filter meteorological characteristic data associated with the power load data, and form a fusion data set with the filtered meteorological characteristic data and power load data; S2: perform data preprocessing on the fused data set; S3: Build a BWO-CNN-BiGRU load forecasting model to perform short-term power load forecasting.

2. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 1 is characterized in that: The method steps for using MIC to screen meteorological characteristic data associated with power load data in S1 are as follows: S11: Calculate the mutual information MI between L and C based on the predicted electric load variable L and weather variable C Where p(l,c) is the joint probability density of L and C, p(l) and p(c) are the marginal probability densities of L and C respectively; S12: Given a pair of integers (x, y), define a grid G(x, y), calculate the mutual information between L and C in the current grid G, and traverse all grids to obtain the maximum MI value for (x, y) f’ MI (D,G(x,y))=maxf MI (D,G(x,y)) In the formula, f MI (D, G(x, y)) is the MI value of the data set D = (L, C) on G(x, y); S13: Normalize the maximum MI value in different grids. The normalized value is between (0,1). The normalization formula is: S14: Take all M G The maximum value of (xy) is taken as the MIC value of the load variable and the weather variable. Where B(n) is a function of n.

3. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 1 is characterized in that: The data preprocessing method for the fused data set in S2 is: missing values ​​in the data set are filled using linear interpolation and the data are normalized.

4. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 1 is characterized in that: The steps for constructing the BWO-CNN-BiGRU load forecasting model in S3 are as follows: S31: Use a sliding window to dynamically take values ​​of the original data, and then construct an input feature map of the CNN feature extraction layer. Use the convolutional neural network CNN to extract a multi-dimensional feature vector that represents the load change, and construct a time series as the sample data of the input layer of the bidirectional gated recurrent unit network prediction model BiGRU; S32: Use the BiGRU model for bidirectional loop training to learn the positive and negative laws of load data. Each GRU layer maintains its own hidden state and updates it according to the current input and the hidden state of the previous time step; S33: The Beluga optimization algorithm (BWO) is used to optimize the model’s weights and other hyperparameters to further improve the prediction model performance, and finally a BWO-CNN-BiGRU load prediction model is constructed.

5. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 4 is characterized in that: The internal hidden layer unit calculation formula of the BiGRU model in S32 is as follows: In the formula, x t represents the state at time t; and Represents the forward and backward states of the hidden layer at time t; and Represents the forward and backward states of the hidden layer at time t; and represents the forward and backward weights of the hidden layer at time t; b t is the bias of the hidden layer state at time t.

6. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 4 is characterized in that: The steps for optimizing the selection by the White Whale Optimization Algorithm BWO in S33 are as follows: S331: Initialization is performed, and the population initialization position is: Where n represents the size of the beluga population; d represents the dimension of the design variable; The fitness values ​​for all beluga whales are: S332: Entering the exploration phase, the mathematical model established by the swimming behavior of the beluga whale, that is, the beluga whale position update formula is as follows: In the formula, T represents the current number of iterations; represents the j-dimensional updated position of the i-th beluga whale, p j is a random number randomly selected from the d-dimensional space, (j=1,2,..d); Indicates that the i-th beluga whale is at p j Position on the dimension; and Represent the current positions of the i-th and r-th beluga whales, respectively. r represents a randomly selected beluga whale. r1 and r2 represent random numbers between (0,1). sin(2πr2) and cos(2πr2) indicate that the mirrored beluga whale fins are facing the water surface. S333: Entering the development stage, the Levy flight strategy was introduced to capture prey. The mathematical model of predation is expressed as: In the formula, T represents the current number of iterations; and They represent the i-th beluga whale at the current position and a randomly selected beluga whale respectively; Indicates the updated position of the i-th beluga whale; represents the optimal position of the white whale; r3 and r4 represent random numbers between (0,1) respectively; C1 represents the random jump intensity of Levy flight, C1=2r4(1-T / T max );L F is Levy's flight function; S334: Entering the whale fall phase, the whale position is updated by using the white whale position and the whale fall step length. The formula is as follows: Where r5, r6 and r7 are random numbers between (0,1); X step represents the step length of the whale's fall, namely: X step =(u b -l b )exp(-C2T / T max ) In the formula, T and T max Respectively represent the current number of iterations and the maximum number of iterations; u b and l b Represent the upper and lower bounds of the search space respectively; C2 represents the step size factor related to the probability of whale falling and the size of the population (C2 = 2W f ×n); the probability of a whale falling is W f is a linear function, that is: W f =0.1-0.5T / T Max where the probability of a whale falling decreases from 0.1 in the initial iteration to 0.05 in the last iteration.

7. The short-term power load forecasting method integrating MIC feature optimization and BWO-CNN-BiGRU according to claim 6 is characterized in that: L F The calculation method is as follows: Where u and v are normally distributed random numbers; β is a constant, β = 1.5.

Images