Short-term load prediction method and device based on variational mode decomposition and recurrent neural network

Abstract

The invention discloses a short-term load prediction method based on variational mode decomposition and a recurrent neural network, and the method specifically comprises the following steps: firstly, obtaining the load data of each charging station through processing the operation order data of the charging station; then, by analyzing the load data of the charging station, constructing a time sequence data set of historical load data of the charging station; secondly, decomposing the load flow data of the charging station based on a variational mode decomposition algorithm; thirdly, training and fitting the load flow data of the charging station based on a recurrent neural network; secondly, adopting a zebra optimization algorithm to adjust and obtain optimal parameters of the variational mode decomposition model and the recurrent neural network; and finally, performing load prediction on each modal component according to the trained recurrent neural network to obtain a load prediction result of the charging station. Noise separation and smooth decomposition processing are carried out on the load data, the load prediction precision of the charging station is improved, and effective services are provided for operation and maintenance of the charging station.

Description

Technical Field

[0001] The present invention belongs to the technical field of charging station load prediction, and more specifically, relates to a short-term load prediction method and device based on variational mode decomposition and recurrent neural network. Background Art

[0002] The development of new energy and equipment upgrades have become the hot topic of the times. With the popularization of electric vehicles, the large-scale construction of charging piles and charging stations, and the access of a large number of high-proportion new energy sources to the power grid, the safety and stability of the power grid have been challenged. In order to solve these problems, power companies have begun to frequently implement peak-shaving and valley-filling schemes and formulate demand response strategies, and the formulation of these strategies depends on the load estimation of the power consumption area. Therefore, short-term power load forecasting of charging stations is crucial to building a safe and stable power grid. Short-term load forecasting usually refers to forecasts within one month, which are mainly used for real-time safety analysis and scheduling of power systems to meet power generation needs and ensure safe and stable operation of the power grid. However, due to external factors, the load changes of charging stations are unstable, and improving the accuracy of prediction has become the main difficulty faced by current short-term load forecasting. Therefore, in view of the problem of unstable changes in power load at charging stations, improving the accuracy of short-term load forecasting at charging stations is of great significance to the safety and maintenance of the power grid. Summary of the invention

[0003] In response to the problem of unstable load changes in charging stations caused by external factors, the present invention proposes a short-term load forecasting method and device based on variational mode decomposition and recurrent neural network, which performs noise separation and smoothing decomposition processing on load data, and uses the decomposed load data to perform load forecasting, thereby solving the problem of unstable load changes in charging stations caused by external factors.

[0004] In order to achieve the above object, the technical solution of the present invention is as follows: a short-term load forecasting method based on variational mode decomposition and recurrent neural network, the method comprising the following steps:

[0005] By processing the charging station operation order data, the historical load data of each charging station is obtained;

[0006] By analyzing the historical load data of charging stations, a time series data set of the historical load data of charging stations is constructed;

[0007] Training and fitting of charging station load flow data based on recurrent neural network;

[0008] The zebra optimization algorithm is used to adjust the optimal parameters of the variational mode decomposition model and the recurrent neural network;

[0009] Based on the adjusted variational mode decomposition model, the charging station load flow data is decomposed to obtain multiple modal components;

[0010] The load prediction of each modal component is performed according to the trained recurrent neural network to obtain the load prediction result of the charging station.

[0011] Preferably, the historical load data of each charging station is obtained by processing the charging station operation order data, and the implementation process is:

[0012] Obtain data from charging station operation orders, including the number of charging piles in use, charging pile number, charging pile start charging time, charging pile end charging time, number of charging orders, etc.;

[0013] Filter the acquired order data and retain the required data, such as the number of charging piles put into use, the proportion of charging piles put into use, the total duration of orders, the number of orders, and the proportion of valid orders;

[0014] By calculating the filtered charging station order data, the historical load data of each charging station is obtained.

[0015] Preferably, the process of analyzing the historical load data of the charging station to construct a time series data set of the historical load data of the charging station is as follows:

[0016] By preprocessing the data, outliers and missing values ​​in the historical load data are removed to ensure the integrity and accuracy of the data. The remaining data are supplemented and calculated to obtain the final charging history data set.

[0017] Considering the load influencing factors such as historical load factors and meteorological factors, the rules for selecting characteristic data are as follows:

[0018] (1) Historical load factor: By analyzing and learning the changing trends and characteristics of historical load data within a specific time period, the model can obtain the law of periodic changes in the charging station load, so that the model can more accurately predict the short-term load of the charging station;

[0019] (2) Meteorological factors: When the outside weather is rainy or other bad weather, residents' travel volume is relatively reduced, and the use of electric vehicles will also decrease, resulting in a relatively reduced load on the charging station on that day; when the outside temperature is high or low, residents are more inclined to choose vehicles as a means of transportation, the use rate of charging vehicles is relatively increased, the number of charging times will increase accordingly, and the total load of the charging station will increase; therefore, weather and temperature are listed as characteristic factors;

[0020] (3) Date factor: On weekdays, residents’ demand for electric vehicles is relatively higher during the morning and evening peak hours compared to other periods; on non-working days, residents have more daytime travel activities and the use rate of electric vehicles is higher;

[0021] According to the above three categories, appropriate feature data are selected to construct feature engineering, and combined with the historical load data of charging stations, a time series data set of the historical load data of charging stations is constructed.

[0022] Preferably, the variational mode decomposition algorithm is used to decompose the charging station load flow data, and the implementation process is as follows:

[0023] S3-1: Construct a variational modal decomposition model for charging station load data. The original data f(t) of the charging station load data is used as input data, and the decomposed modal components u k (t) is the output data. The bandwidth description of each mode is obtained by demodulating the charging station load flow data signal through Gaussian smoothing. The variational mode decomposition model construction formula is as follows:

[0024]

[0025] Where f(t) is the input data; u k (t) is the modal component after decomposition; t is the current time; is the signal ladder; w k is the center frequency of the kth mode; δ(t) is the unit impulse signal, is the square root of the L2 norm gradient; e, π are mathematical constants; j is the symbol of the imaginary number;

[0026] Solve the variational modal decomposition model of power load data. The solution of the variational problem can be regarded as the conditional maximum of a multivariate function. The constrained variational problem is transformed into an unconstrained variational problem using the Lagrange multiplier method. The transformed Lagrangian function L({u k},{ω k},λ) is:

[0027]

[0028] Where f(t) is the input data; u k (t) is the modal component after decomposition; w k is the kth mode center frequency; α is the quadratic penalty factor; λ is the Lagrangian factor.

[0029] The alternating multiplier method is used to search for the Lagrangian function L({u k},{ω k},λ), the constraint parameters are continuously updated, and the update formula of the constraint parameters is as follows:

[0030]

[0031] Where f(w) is the Fourier transform of the input data, i.e., the frequency domain representation; is the frequency domain representation of the current i-th modal component; is the kth modal component after update; is the central frequency domain of the current kth modal component; is the updated central frequency domain of the kth modal component; is the frequency domain representation of the current Lagrangian factor; is the updated Lagrangian factor; K is the number of modal components; η is the noise tolerance, which is used to buffer the interference caused by the presence of strong noise.

[0032] Repeat the update of parameter values ​​until the following constraints are met, stop the iteration, and obtain the optimal modal component of the charging station load decomposition:

[0033]

[0034] In the formula, is the kth modal component before updating; is the kth modal component after updating; ε is the determination accuracy; K is the number of modal components.

[0035] According to the periodic characteristics of the modal components, they are divided into periodic modal components and non-periodic modal components.

[0036] Preferably, the training and fitting of the charging station load flow data based on the recurrent neural network is implemented as follows:

[0037] A gated recurrent neural network is selected to train and fit the load flow data of charging stations;

[0038] The gated recurrent neural network structure is composed of three gated units for controlling information transmission, namely, an update gate, a reset gate, and an output gate. The update gate is used to indicate the degree of influence of the state information at the previous moment on the current state, and the reset gate is used to determine the degree of discarding the state information at the previous moment. The overall framework is as follows:

[0039] Update Gate: Z t =σ(W Z ·[x t ,h t-1 ]) (8)

[0040] Reset gate: r t =σ(W r ·[x t ,h t-1 ]) (9)

[0041]

[0042] Output gate: o t =σ(W o ·h t ) (12)

[0043] In the formula, x t Indicates the current charging station load data; o t represents the predicted output of the charging station operation data at the current moment; h t represents the hidden state of the gated recurrent neural network updated at the current moment; h t-1 Represents the hidden state of the gated recurrent neural network at the previous moment; Represents the hidden state of the candidate gated recurrent neural network at the current moment; Z t represents the output of the update gate; r t represents the output of the reset gate; tanh represents the hyperbolic tangent activation function; σ represents the nonlinear activation function; W Z represents the update gate weight matrix; W r Represents the reset gate weight matrix; W h represents the candidate hidden layer weight matrix, W O represents the output layer weight matrix; E represents the identity matrix.

[0044] (1) Reset gate: determines the degree to which the state information of the previous moment is discarded. t The value range is [0,1], r t The smaller it is, the more state information is discarded. The reset gate helps capture short-term dependencies in the time series by discarding state information that is irrelevant to the charging station load forecast.

[0045] (2) Update gate: indicates the influence of the state information at the previous moment on the current state. The update gate has the function of a forget gate, which selectively forgets the state information at the previous moment. t-1 , and will selectively memorize the current state information, merge the two and update the current state information h t .

[0046] Construct a bidirectional gated recurrent neural network model. The bidirectional gated recurrent neural network is composed of a forward gated recurrent neural network and a backward gated recurrent neural network. The input sequence of the forward gated recurrent neural network is the original sequence {L(0), L(1), ..., L(n)}, and n output vectors are generated through the neural network: GRU L ={h L (0),...,h L The input sequence of the backward gated recurrent neural network is the reverse sequence {L(n), L(n-1), ..., L(0)}, and the neural network generates n output vectors: GRU R ={h′ L (n),...,h′ L(0)}, therefore, the power load prediction result of the bidirectional gated recurrent neural network model at time t is:

[0047] h t =βh t +γh t +b (13)

[0048] O t =σ(W B ·h t ) (14)

[0049] In the formula, h t represents the state value generated by the original sequence through the bidirectional gated recurrent neural network; h t represents the state value generated by the reverse sequence through the bidirectional gated recurrent neural network; h t represents the hidden state of the bidirectional gated recurrent neural network updated at the current moment; β, γ represent the output weights; b represents the offset; O t W represents the predicted load of the bidirectional gated recurrent neural network model at time t; B Represents the weight matrix of the output layer of the bidirectional gated recurrent neural network.

[0050] Preferably, the zebra optimization algorithm is used to adjust and obtain the optimal parameters of the variational mode decomposition model and the recurrent neural network, and the implementation process is:

[0051] The zebra population is initialized using random numbers. The initialization formula of the zebra population is as follows:

[0052] x i,v =lb v +q·(ub v -lb v ) (15)

[0053] In the formula, x i,v represents the initialization position of the vth dimension of the i-th zebra; ub v Indicates the upper boundary of optimization; lb v represents the lower boundary of optimization; q represents a random number between 0 and 1;

[0054] By simulating the foraging behavior of zebras, the positions of population members are updated to find the optimal solution. The formula for updating the position information of the zebra population during foraging is:

[0055]

[0056] In the formula, Indicates the updated position of the i-th zebra in the foraging phase; represents the new position of the vth dimension after the ith zebra update; x i,vIndicates the current v-th dimension position of the i-th zebra; F i new,P1 represents the fitness value after the i zebras are updated; X i Indicates the current position of the i-th zebra; F i Indicates the current fitness value of the i-th zebra; PZ v represents the position of the vth dimension of the pioneer zebra; q represents a random number between 0 and 1; I represents a random value belonging to the set {1,2}.

[0057] Simulate the defensive actions of zebras to prevent attacks by predators to update the positions of population members. When facing attacks, zebras will adopt two strategies. When the predator is a lion, the zebra will adopt a defensive strategy of running away to escape the lion's attack; when the predator is another animal, the zebra will adopt a defensive strategy of counterattack. The zebras in the population will move to the position of the attacked zebra and adopt certain strategies to confuse the predator. The formula for updating the zebra position information is:

[0058]

[0059] In the formula, Indicates the position of the i-th zebra after update in the defense phase; represents the new position of the vth dimension after the ith zebra update; x i,v Indicates the current v-th dimension position of the i-th zebra; F i new,P2 represents the fitness value of the i-th zebra after update; X i Indicates the current position of the i-th zebra; F i represents the current fitness value of the i-th zebra; AZ represents the position of the attacked zebra, AZ v represents the v-th dimension position of the zebra; P s represents the switching probability of the two strategies of the zebra population; m represents the current number of iterations; M represents the maximum number of iterations; q represents a random number between 0 and 1; R is a constant equal to 0.01; I represents a random value belonging to the set {1,2}.

[0060] Preferably, the load prediction of each modal component is performed according to the trained recurrent neural network, and finally the load prediction of the charging station is realized, and the realization process is:

[0061] The periodic modal components are directly based on the historical load data, and the load forecasting of each modal component is carried out through a bidirectional gated recurrent neural network;

[0062] The non-periodic modal components are processed by the middle layer and the characteristic data of the load data are added to the bidirectional gated recurrent neural network for load forecasting.

[0063] The load forecast results of each modal component are reconstructed to obtain the load forecast result of the charging station. The calculation formula of the load forecast result of the charging station is as follows:

[0064]

[0065] In the formula, h a (t) represents the load forecast result of the charging station; h i (t) represents the load prediction result of the i-th modal component; K represents the number of modal components.

[0066] A short-term load forecasting device based on variational mode decomposition and recurrent neural network is implemented as follows:

[0067] A data acquisition module, used to acquire historical load data of the charging station;

[0068] A data preprocessing module, used to process the charging station data information, determine the characteristic vector of the data influencing factors, and generate a time series data set of the charging station historical load data;

[0069] Parameter optimization module, used to adjust the variational mode decomposition model and recurrent neural network model based on the zebra optimization algorithm to obtain the optimal parameters of the model;

[0070] A load decomposition module is used to decompose the load flow data of the charging station based on the variational mode decomposition algorithm to obtain multiple modal components;

[0071] The load prediction module is used to perform load prediction on each modal component based on a recurrent neural network to realize load prediction of the charging station.

[0072] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the steps of the short-term load forecasting method based on variational mode decomposition and recurrent neural network.

[0073] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the short-term load forecasting method based on variational mode decomposition and recurrent neural network is implemented.

[0074] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0075] The technical solution of the present invention, first, by processing the charging station operation order data, obtain the load data of each charging station; then, by analyzing the charging station load data, construct a time series data set of the charging station historical load data; then, based on the variational mode decomposition algorithm, the charging station load flow data is decomposed; again, based on the recurrent neural network, the charging station load flow data is trained and fitted; secondly, the zebra optimization algorithm is used to adjust the optimal parameters of the variational mode decomposition model and the recurrent neural network; finally, the load prediction of each modal component is performed according to the trained recurrent neural network, and finally the load prediction of the charging station is realized. The load data is decomposed into multiple modal components by the variational mode decomposition algorithm, the dynamic characteristics in the load signal are effectively captured, and the load prediction is performed for the decomposed modal variables, so as to further improve the load prediction accuracy of the charging station. The method of the present invention can separate the noise and smooth decomposition of the load data and conduct targeted analysis on the decomposed load data, effectively ensuring that an accurate data basis is provided for the power company to dispatch the power load of the charging station, and providing effective services for the operation and maintenance of the charging station. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flow chart of a short-term load forecasting method based on variational mode decomposition and recurrent neural network provided by an embodiment of the present invention;

[0077] Figure 2 A short-term load forecasting device based on variational mode decomposition and recurrent neural network is provided in an embodiment of the present invention;

[0078] Figure 3 It is a short-term load prediction result of a charging station based on the method of the present invention provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0080] Example 1: Figure 1 A short-term load forecasting method based on variational mode decomposition and recurrent neural network according to an example of the present invention is shown, and the method comprises the following steps:

[0081] Step 1: Obtain the load data of each charging station by processing the charging station operation order data;

[0082] Step 2: Analyze the charging station load data to build a time series data set of the charging station historical load data;

[0083] Step 3: Decompose the charging station load flow data based on the variational mode decomposition model;

[0084] Step 4: Train and fit the charging station load flow data based on the recurrent neural network;

[0085] Step 5: Use the zebra optimization algorithm to adjust the optimal parameters of the variational mode decomposition model and the recurrent neural network;

[0086] Step 6: Perform load forecasting on each modal component based on the trained recurrent neural network to obtain the load forecasting result of the charging station.

[0087] Preferably, in step S1, the load data of each charging station is obtained by processing the charging station operation order data, and the implementation process is:

[0088] S1-1: Obtain data from charging station operation orders, including the number of charging piles put into use, charging pile number, charging pile start charging time, charging pile end charging time, number of charging orders, etc.;

[0089] S1-2: Filter the acquired order data and retain the required data, such as the number of charging piles put into use, the proportion of charging piles put into use, the total duration of orders, the number of orders, and the proportion of valid orders;

[0090] S1-3: By calculating the filtered charging station order data, the historical load data of each charging station is obtained.

[0091] Preferably, in step S2, the process of analyzing the charging station load data to construct a time series data set of the charging station historical load data is as follows:

[0092] S2-1: By preprocessing the data, outliers and missing values ​​in the historical load data are removed to ensure the integrity and accuracy of the data. The remaining data are supplemented and calculated to obtain the final charging history data set.

[0093] S2-2: Considering the load influencing factors such as historical load factors and meteorological factors, the selected features are as follows:

[0094] (1) Historical load factor: By analyzing and learning the changing trends and characteristics of historical load data within a specific time period, the model can obtain the law of periodic changes in the charging station load, so that the model can more accurately predict the short-term load of the charging station;

[0095] (2) Meteorological factors: When the outside weather is rainy or other bad weather, residents' travel volume is relatively reduced, and the use of electric vehicles will also decrease, resulting in a relatively reduced load on the charging station on that day; when the outside temperature is high or low, residents are more inclined to choose vehicles as a means of transportation, the use rate of charging vehicles is relatively increased, the number of charging times will increase accordingly, and the total load of the charging station will increase; therefore, weather and temperature are listed as characteristic factors;

[0096] (3) Date factor: On weekdays, residents’ demand for electric vehicles is relatively higher during the morning and evening peak hours compared to other periods; on non-working days, residents have more daytime travel activities and the use rate of electric vehicles is higher;

[0097] S2-3: Based on the above three categories, appropriate features are selected to construct feature engineering, and combined with the load data of the charging station, a time series dataset of the historical load data of the charging station is constructed.

[0098] Preferably, in step S3, the variational mode decomposition algorithm is used to decompose the charging station load flow data, and the implementation process is as follows:

[0099] S3-1: Construct a variational modal decomposition model for charging station load data. The original data f(t) of the charging station load data is used as input data, and the decomposed modal components u k (t) is the output data. The bandwidth description of each mode is obtained by demodulating the charging station load flow data signal through Gaussian smoothing. The variational mode decomposition algorithm is constructed as follows:

[0100]

[0101] Where f(t) is the input data; u k (t) is the modal component after decomposition; t is the current time; is the signal ladder; w k is the center frequency of the kth mode; δ(t) is the unit impulse signal, is the square root of the L2 norm gradient; e and π are mathematical constants; j is the imaginary number symbol.

[0102] S3-2: Solve the variational modal decomposition model of power load data. The solution of the variational problem can be regarded as the conditional minimum of a multivariate function. The constrained variational problem is transformed into an unconstrained variational problem using the Lagrange multiplier method. The transformed Lagrangian function L({u k},{ω k},λ) is:

[0103]

[0104] Where f(t) is the input data; u k(t) is the modal component after decomposition; w k is the kth mode center frequency; α is the quadratic penalty factor; λ is the Lagrangian factor.

[0105] S3-3: Searching for the Lagrangian function L({u k},{ω k},λ), and continuously update the constraint parameters. The update formulas of the three parameters are as follows:

[0106]

[0107] Where f(w) is the Fourier transform of the input data, i.e., the frequency domain representation; is the frequency domain representation of the current i-th modal component; is the kth modal component after update; is the central frequency domain of the current kth modal component; is the center frequency domain of the kth modal component after update; is the frequency domain representation of the current Lagrangian factor; is the updated Lagrangian factor; K is the number of modal components; η is the noise tolerance, which is used to buffer the interference caused by the presence of strong noise.

[0108] S3-4: Repeat the update of parameter values ​​until the following constraints are met, stop the iteration, and obtain the optimal modal component of the charging station load decomposition:

[0109]

[0110] In the formula, is the kth modal component before updating; is the kth modal component after updating; ε is the determination accuracy; K is the number of modal components.

[0111] S3-5: According to the periodic characteristics of the modal components, they are divided into non-periodic principal components and multiple smooth periodic components.

[0112] Preferably, in step S4, the training and fitting of the charging station load flow data based on the recurrent neural network is implemented as follows:

[0113] S4-1: Select a recurrent neural network model. Select a gated recurrent neural network to train and fit the charging station load flow data. The gated recurrent neural network model can capture the dependencies with large time step distances in the time series.

[0114] S4-2: Construct a gated recurrent neural network model. The gated recurrent neural network structure is mainly composed of three gated units that control information transmission: update gate, reset gate, and output gate. Its overall framework is as follows:

[0115] Update Gate: Z t =σ(W Z ·[x t ,h t-1 ]) (8)

[0116] Reset gate: r t =σ(W r ·[x t ,h t-1 ]) (9)

[0117]

[0118] Output gate: o t =σ(W o ·h t ) (12)

[0119] In the formula, x t Indicates the current charging station load data; o t represents the predicted output of the charging station operation data at the current moment; h t represents the hidden state of the gated recurrent neural network updated at the current moment; h t-1 Represents the hidden state of the gated recurrent neural network at the previous moment; Represents the hidden state of the candidate gated recurrent neural network at the current moment; Z t represents the output of the update gate; r t represents the output of the reset gate; tanh represents the hyperbolic tangent activation function; σ represents the nonlinear activation function; W Z represents the update gate weight matrix; W r Represents the reset gate weight matrix; W h represents the candidate hidden layer weight matrix, W O represents the output layer weight matrix; E represents the identity matrix.

[0120] (1) Reset gate: determines the degree to which the state information of the previous moment is discarded. t The value range is [0,1], r t The smaller it is, the more state information is discarded. The reset gate helps capture short-term dependencies in the time series by discarding state information that is irrelevant to the charging station load forecast.

[0121] (2) Update gate: indicates the influence of the state information at the previous moment on the current state. The update gate has the function of a forget gate, which selectively forgets the state information at the previous moment. t-1 , and will selectively memorize the current state information, merge the two and update the current state information h t .

[0122] S4-3: Construct a bidirectional gated recurrent neural network model. The bidirectional gated recurrent neural network is composed of a forward gated recurrent neural network and a backward gated recurrent neural network. The input sequence of the forward gated recurrent neural network is the original sequence {L(0), L(1), ..., L(n)}, and n output vectors are generated through the neural network: GRU L ={h L (0),...,h L The input sequence of the backward gated recurrent neural network is the reverse sequence {L(n), L(n-1), ..., L(0)}, and the neural network generates n output vectors: GRU R ={h′ L (n),...,h′ L (0)}, therefore, the power load prediction result of the bidirectional gated recurrent neural network model at time t is:

[0123] h t =βh t +γh t +b (13)

[0124] O t =σ(W B ·h t ) (14)

[0125] In the formula, h t represents the state value generated by the original sequence through the bidirectional gated recurrent neural network; h t represents the state value generated by the reverse sequence through the bidirectional gated recurrent neural network; h t represents the hidden state of the bidirectional gated recurrent neural network updated at the current moment; β, γ represent the output weights; b represents the offset; O t W represents the predicted load of the bidirectional gated recurrent neural network model at time t; B Represents the weight matrix of the output layer of the bidirectional gated recurrent neural network.

[0126] Preferably, in step S5, the zebra optimization algorithm is used to adjust and obtain the optimal parameters of the variational mode decomposition model and the recurrent neural network, and the implementation process is:

[0127] S5-1: Initialize the zebra population. Use random numbers to initialize the zebra population:

[0128] x i,v =lb v +q·(ub v -lb v ) (15)

[0129] In the formula, x i,vrepresents the initialization position of the vth dimension of the i-th zebra; ub v Indicates the upper boundary of optimization; lb v represents the lower boundary of optimization; q represents a random number between 0 and 1.

[0130] S5-2: Foraging behavior. This behavior is to update the positions of population members by simulating the foraging behavior of zebras to find the optimal solution. In the zebra optimization algorithm, the best zebras in the population are considered pioneer zebras. They will actively look for high-quality food and lead other members of the population to move to the location of the food. The formula for updating the position information of the zebra population during the foraging process is:

[0131]

[0132] In the formula, Indicates the updated position of the i-th zebra in the foraging phase; represents the new position of the vth dimension after the ith zebra update; x i,v Indicates the current v-th dimension position of the i-th zebra; F i new,P1 represents the fitness value after the i zebras are updated; X i Indicates the current position of the i-th zebra; F i Indicates the current fitness value of the i-th zebra; PZ v represents the position of the vth dimension of the pioneer zebra; q represents a random number between 0 and 1; I represents a random value belonging to the set {1,2}.

[0133] S5-3: Defensive behavior. This behavior simulates the defensive actions of zebras to prevent attacks from predators, so as to update the positions of population members. When facing attacks, zebras will adopt two strategies. When the predator is a lion, the zebra will adopt a defensive strategy of running away to escape the lion's attack; when the predator is another animal, the zebra will adopt a defensive strategy of counterattack. The zebras in the population will move to the position of the attacked zebra and adopt certain strategies to confuse the predator. The formula for updating the zebra position information is:

[0134]

[0135] In the formula, Indicates the position of the i-th zebra after update in the defense phase; represents the new position of the vth dimension after the ith zebra update; x i,v Indicates the current v-th dimension position of the i-th zebra; F i new,P2 represents the fitness value of the i-th zebra after update; X i Indicates the current position of the i-th zebra; F i represents the current fitness value of the i-th zebra; AZ represents the position of the attacked zebra, AZv represents the v-th dimension position of the zebra; P s represents the switching probability of the two strategies of the zebra population; m represents the current number of iterations; M represents the maximum number of iterations; q represents a random number between 0 and 1; R is a constant equal to 0.01; I represents a random value belonging to the set {1,2}.

[0136] Preferably, in step S6, the load prediction of each modal component is performed according to the trained recurrent neural network, and finally the load prediction of the charging station is realized, and the realization process is:

[0137] S6-1: The periodic modal components are directly based on the historical load data, and the load forecasting of each modal component is carried out through a bidirectional gated recurrent neural network;

[0138] S6-2: The non-periodic modal components are processed by the intermediate layer and the characteristic data of the load data are added to the bidirectional gated recurrent neural network for load forecasting;

[0139] S6-3: Reconstruct the load prediction results of each modal component to obtain the load of the charging station:

[0140]

[0141] In the formula, h a (t) represents the predicted load of the charging station; h i (t) represents the predicted load of the i-th modal component; K represents the number of modal components.

[0142] Example 2: Figure 2 As shown, an embodiment of the present invention provides a short-term load forecasting device based on variational mode decomposition and recurrent neural network, and the device includes the following modules:

[0143] A data acquisition module, used to acquire historical load data of the charging station;

[0144] A data preprocessing module, used to process the charging station data information, determine the characteristic vector of the data influencing factors, and generate a time series data set of the charging station historical load data;

[0145] Parameter optimization module, used to adjust the variational mode decomposition model and recurrent neural network model based on the zebra optimization algorithm to obtain the optimal parameters of the model;

[0146] A load decomposition module is used to decompose the load flow data of the charging station based on the variational mode decomposition algorithm to obtain multiple modal components;

[0147] The load prediction module is used to perform load prediction on each modal component based on a recurrent neural network to realize load prediction of the charging station.

[0148] The specific implementation of each module can refer to the description of the above method embodiment, and the embodiment of the present invention will not be repeated.

[0149] like Figure 3 As shown in FIG. 1 , a short-term load forecasting result of a charging station based on the method of the present invention is provided in an embodiment of the present invention. The forecasting result is specifically analyzed as follows:

[0150] Depend on Figure 3 It can be seen that the prediction results of the method provided by the present invention are almost consistent with the true value. The method of the present invention is evaluated using the mean absolute percentage error and the coefficient of determination as evaluation indicators. The example results of the present invention are as follows: the mean absolute percentage error is 0.052, and the coefficient of determination is 0.9978. It can be seen that the short-term load prediction method for charging stations provided by the present invention has achieved very good prediction results, which fully proves the accuracy and practicality of the present invention.

[0151] Embodiment 3: A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the short-term load forecasting method based on variational mode decomposition and recurrent neural network.

[0152] Embodiment 4: A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the short-term load forecasting method based on variational mode decomposition and recurrent neural network when executing the program.

[0153] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A short-term load forecasting method based on variational mode decomposition and recurrent neural network, characterized by: The method comprises the following steps: By processing the charging station operation order data, the historical load data of each charging station is obtained; By analyzing the historical load data of charging stations, a time series data set of the historical load data of charging stations is constructed; Training and fitting of charging station load flow data based on recurrent neural network; The zebra optimization algorithm is used to adjust the optimal parameters of the variational mode decomposition model and the recurrent neural network; Based on the adjusted variational mode decomposition model, the charging station load flow data is decomposed to obtain multiple modal components; The load prediction of each modal component is performed according to the trained recurrent neural network to obtain the load prediction result of the charging station.

2. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The process of obtaining the historical load data of each charging station by processing the charging station operation order data is as follows: Obtain order data from charging station operation orders; Filter the acquired order data and retain the required data; By calculating the filtered charging station order data, the historical load data of each charging station is obtained.

3. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The process of constructing a time series data set of the historical load data of the charging station by analyzing the historical load data of the charging station is as follows: By preprocessing the historical load data, removing the outliers and missing values ​​in the historical load data, supplementing and calculating the remaining data, the final charging history data set is obtained; Feature selection is performed based on historical load factors, meteorological factors and date factors to obtain feature data; Feature engineering is constructed based on the feature data, and a time series data set of the historical load data of the charging station is constructed in combination with the historical load data of the charging station.

4. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The process of decomposing the charging station load flow data based on the variational mode decomposition algorithm is as follows: Construct a variational modal decomposition model for charging station load data, taking the original data f(t) of the charging station load data as input data, and decomposing each modal component u k (t) is the output data. The bandwidth description of each modal component is obtained by demodulating the charging station load flow data signal through Gaussian smoothing. The variational modal decomposition model construction formula is as follows: Where f(t) is the input data; u k (t) is the modal component after decomposition; t is the current time; is the signal ladder; w k is the center frequency of the kth mode; δ(t) is the unit impulse signal, is the square root of the L2 norm gradient; e, π are mathematical constants; j is the symbol of the imaginary number; Solve the variational modal decomposition model of power load data. The solution of the variational problem is to obtain the conditional maximum value of the multivariate function. The constrained variational problem is transformed into an unconstrained variational problem using the Lagrange multiplier method. The transformed Lagrangian function L({u k },{ω k },λ) is: Where f(t) is the input data; u k (t) is the modal component after decomposition; w k is the kth mode center frequency; α is the quadratic penalty factor; λ is the Lagrangian factor; The alternating multiplier method is used to search for the Lagrangian function L({u k },{ω k },λ), the constraint parameters are continuously updated, and the update formula of the constraint parameters is as follows: Where f(w) is the Fourier transform of the input data, i.e., the frequency domain representation; is the frequency domain representation of the current i-th modal component; is the kth modal component after update; is the central frequency domain of the current kth modal component; is the updated central frequency domain of the kth modal component; is the frequency domain representation of the current Lagrangian factor; is the updated Lagrangian factor; K is the number of modal components; η is the noise margin, which is used to buffer the interference caused by the presence of strong noise; Repeatedly update the values ​​of the constraint parameters until the following constraint conditions are met, stop the iteration, and obtain the optimal modal component of the charging station load flow data decomposition: In the formula, is the kth modal component before updating; is the kth modal component after updating; ε is the determination accuracy; K is the number of modal components; According to the periodic characteristics of the modal components, they are divided into periodic modal components and non-periodic modal components.

5. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The process of training and fitting the charging station load flow data based on the recurrent neural network is as follows: A gated recurrent neural network is selected to train and fit the load flow data of charging stations; The gated recurrent neural network structure is composed of three gated units for controlling information transmission, namely, an update gate, a reset gate, and an output gate. The update gate is used to indicate the degree of influence of the state information at the previous moment on the current state, and the reset gate is used to determine the degree of discarding the state information at the previous moment. The overall framework is as follows: Update Gate: Z t =σ(W Z ·[x t ,h t-1 ]) (8) Reset gate: r t =σ(W r ·[x t ,h t-1 ]) (9) Output gate: o t =σ(W o ·h t ) (12) In the formula, x t Indicates the current charging station load data; o t represents the predicted output of the charging station operation data at the current moment; h t represents the hidden state of the gated recurrent neural network updated at the current moment; h t-1 Represents the hidden state of the gated recurrent neural network at the previous moment; Represents the hidden state of the candidate gated recurrent neural network at the current moment; Z t represents the output of the update gate; r t represents the output of the reset gate; tanh represents the hyperbolic tangent activation function; σ represents the nonlinear activation function; W Z represents the update gate weight matrix; W r Represents the reset gate weight matrix; W h represents the candidate hidden layer weight matrix, W O represents the output layer weight matrix; E represents the identity matrix; A bidirectional gated recurrent neural network model is constructed. The bidirectional gated recurrent neural network is composed of a forward gated recurrent neural network and a backward gated recurrent neural network. The input sequence of the forward gated recurrent neural network is the original sequence {L(0), L(1), ..., L(n)}, and n output vectors are generated through the neural network: GRU L ={h L (0),...,h L (n)}, the input sequence of the backward gated recurrent neural network is the reverse sequence {L(n), L(n-1), ..., L(0)}, and n output vectors are generated through the neural network: GRU R ={h′ L (n),...,h′ L (0)}, the power load prediction result of the bidirectional gated recurrent neural network model at time t is: h t =βh t +γh t +b (13) The t =σ(W B ·h t ) (14) In the formula, h t represents the state value generated by the original sequence through the bidirectional gated recurrent neural network; h t represents the state value generated by the reverse sequence through the bidirectional gated recurrent neural network; h t represents the hidden state of the bidirectional gated recurrent neural network updated at the current moment; β, γ represent the output weights; b represents the offset; O t W represents the predicted load of the bidirectional gated recurrent neural network model at time t; B Represents the weight matrix of the output layer of the bidirectional gated recurrent neural network.

6. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The zebra optimization algorithm is used to adjust the optimal parameters of the variational mode decomposition model and the recurrent neural network, and the process is as follows: The zebra population is initialized using random numbers. The initialization formula of the zebra population is as follows: x i,v =lb v +q·(ub v -lb v ) (15) In the formula, x i,v represents the initialization position of the vth dimension of the i-th zebra; ub v Indicates the upper boundary of optimization; lb v represents the lower boundary of optimization; q represents a random number between 0 and 1; By simulating the foraging behavior of zebras, the positions of population members are updated to find the optimal solution. The formula for updating the position information of the zebra population during foraging is: In the formula, Indicates the updated position of the i-th zebra in the foraging phase; represents the new position of the vth dimension after the ith zebra update; x i,v Indicates the current v-th dimension position of the i-th zebra; F i new,P1 represents the fitness value after the i zebras are updated; X i Indicates the current position of the i-th zebra; F i Indicates the current fitness value of the i-th zebra; PZ v represents the position of the vth dimension of the pioneer zebra; q represents a random number between 0 and 1; I represents a random value belonging to the set {1,2}; The defensive actions of zebras to prevent attacks by predators are simulated to update the positions of population members. The formula for updating the zebra position information is: In the formula, Indicates the position of the i-th zebra after update in the defense phase; represents the new position of the vth dimension after the ith zebra update; x i,v Indicates the current v-th dimension position of the i-th zebra; F i new,P2 represents the fitness value of the i-th zebra after update; X i Indicates the current position of the i-th zebra; F i represents the current fitness value of the i-th zebra; AZ represents the position of the attacked zebra, AZ v represents the v-th dimension position of the zebra; P s represents the switching probability of the two strategies of the zebra population; m represents the current number of iterations; M represents the maximum number of iterations; q represents a random number between 0 and 1; R is a constant equal to 0.01; I represents a random value belonging to the set {1,2}.

7. The short-term load forecasting method based on variational mode decomposition and recurrent neural network according to claim 1 is characterized in that: The load prediction of each modal component is performed according to the trained recurrent neural network to obtain the load prediction result of the charging station. The process is as follows: The periodic modal components are directly based on the historical load data, and the load forecasting of each modal component is carried out through a bidirectional gated recurrent neural network; The non-periodic modal components are processed by the middle layer and added with the characteristic data of the historical load data to be brought into the bidirectional gated recurrent neural network for load forecasting; The load forecast results of each modal component are reconstructed to obtain the load forecast result of the charging station. The calculation formula of the load forecast result of the charging station is as follows: In the formula, h a (t) represents the load forecast result of the charging station; h i (t) represents the load prediction result of the i-th modal component; K represents the number of modal components.

8. A short-term load forecasting device based on variational mode decomposition and recurrent neural network, characterized in that: The device comprises: A data acquisition module, used to acquire historical load data of the charging station; A data preprocessing module, used to process the charging station data information, determine the characteristic vector of the data influencing factors, and generate a time series data set of the charging station historical load data; Parameter optimization module, used to adjust the variational mode decomposition model and recurrent neural network model based on the zebra optimization algorithm to obtain the optimal parameters of the model; A load decomposition module is used to decompose the load flow data of the charging station based on the variational mode decomposition algorithm to obtain multiple modal components; The load prediction module is used to perform load prediction on each modal component based on a recurrent neural network to realize load prediction of the charging station.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the steps of the short-term load forecasting method based on variational mode decomposition and recurrent neural network as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the short-term load forecasting method based on variational mode decomposition and recurrent neural network as described in any one of claims 1 to 7 above is implemented.

Images