A method, device and medium for predicting charging load of electric vehicle based on NAHL

Abstract

The present invention discloses a method, device and medium for predicting charging load of electric vehicles based on NAHL, the method comprising: optimizing VMD parameters using an improved RSEA algorithm, decomposing and preprocessing charging load data of electric vehicles, constructing independent neural network models with enhanced hidden layers for each component obtained; for each model, selecting RMSE and SMAPE as optimization indicators, using an improved RSEA algorithm to perform multi-objective optimization and training, selecting the optimal solution closest to the center point in the Pareto front as a compromise solution, and finally outputting the optimal parameters and prediction results of the NAHL neural network model, aggregating the prediction values ​​of each NAHL model, and outputting the final prediction result. The present invention adopts a new artificial neural network model, uses a multi-objective optimization method to optimize the neural network parameters, can simultaneously focus on the accuracy of the prediction model and the overall difference between the predicted value and the actual value, and obtain more accurate prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load forecasting, and in particular to a method, a device and a medium for forecasting charging load of an electric vehicle based on NAHL. Background Art

[0002] With the technological development of the electric vehicle industry, the continuous improvement of battery and vehicle technology, the large-scale application of electric vehicles is expanding, and electric vehicles have also attracted much attention for their pollution-free and energy-saving and emission-reduction capabilities. However, due to the strong intermittent and random nature of electric vehicle charging, the large-scale access of electric vehicles to the power grid will lead to increased peak loads and increased energy consumption, which will affect the planning and operation of the power system. In this context, characteristic analysis and prediction of electric vehicle charging loads can better plan and manage the use of charging facilities and improve the utilization rate of charging facilities; it can better dispatch power resources and help users better arrange charging plans; therefore, it is particularly important to predict electric vehicle charging loads.

[0003] The basis for solving the above problems is to accurately predict the charging load of electric vehicles, so it is urgent to establish an effective load forecasting model. Traditional machine learning models have limited learning capabilities for complex nonlinear sequences, and most of them use single-objective algorithms. They focus on improving the accuracy of the prediction model, but rarely pay attention to the stability of the prediction results. They are not suitable for scenarios where the relative degree of error needs to be paid attention to. Therefore, a new method is needed to predict the charging load of electric vehicles. Summary of the invention

[0004] Purpose of the invention: The present invention provides a method, device and medium for predicting electric vehicle charging load based on NAHL, which uses an improved evolutionary algorithm based on radial space partitioning to perform multi-objective optimization on NAHL model parameters to better predict electric vehicle charging load.

[0005] Technical solution: The NAHL-based electric vehicle charging load prediction method of the present invention comprises the following steps:

[0006] (1) Preprocessing the historical charging load data of electric vehicles and dividing it into a training data set and a test data set;

[0007] (2) VMD is used to decompose and preprocess the charging load data of electric vehicles. The improved evolutionary algorithm based on radial space partitioning, namely, improved RSEA, is used to optimize the parameters of VMD, obtain K components and construct the corresponding K input matrices, and determine the training matrix and test matrix of each component;

[0008] (3) Construct a neural network NAHL model with enhanced hidden layers;

[0009] (4) The root mean square error and symmetric mean absolute percentage error between the predicted value and the actual value of the electric vehicle charging load are used as the fitness function; and the training matrix and test matrix of each component are respectively introduced into the established NAHL model;

[0010] (5) The improved RSEA algorithm is used to optimize and train the NAHL model, select the compromise solution of the Pareto frontier, and output the optimal parameters of the NAHL neural network sub-model;

[0011] (6) Calculate the prediction value of each NAHL model using the trained NAHL model and its corresponding test matrix in step (2);

[0012] (7) Aggregate the prediction values ​​of each NAHL model and output the final prediction result.

[0013] Furthermore, the preprocessing in step (1) is to fill in the empty values ​​in the data:

[0014] y k =(y k-1 +y k+1 ) / 2

[0015] Among them, y k is the filling data of the missing time, y k-1 is the data at the same time in the previous cycle, y k+1 The data is the same time in the next cycle.

[0016] Furthermore, the improved evolutionary algorithm based on radial space partitioning in step (2) is:

[0017] Use Tent-Logistic-Cosine chaotic map initialization, the initialization strategy is as follows:

[0018]

[0019] Among them, x(i) is the current state, x(i+1) is the next state, r is the mapping parameter, r∈[0,1]; then through PF i =PF min +x i * (PF max -PF min ) to obtain the initial result; PF max is the upper limit of the i-th decision variable, PF min is the lower limit of the ith decision variable, x i is the value initialized by the Tent-Logistic-Cosine chaotic map, PF irepresents the initialization result of the i-th decision variable; the mutation operation is improved, and an exponentially decreasing adaptive mutation probability strategy is proposed. The exponentially decreasing adaptive mutation probability strategy is as follows:

[0020] MP=MP min *(MP max -MP min )^[1 / (1+K*FE / FE max )]

[0021] In the formula, K is the parameter that controls the exponential decrease. The larger K is, the faster the decrease speed is. max is the maximum mutation probability, MP min is the minimum mutation probability, FE max is the maximum number of iterations, FE is the current number of iterations; in the early stage of iteration, the mutation rate is high, and widely distributed solutions are generated, which improves the diversity of the algorithm; in the middle stage of iteration, the mutation probability will gradually decrease; in the late stage of iteration, the mutation probability is the smallest, and the solutions obtained are more dense, which improves the convergence of the algorithm.

[0022] Furthermore, the implementation process of step (2) is as follows:

[0023] (21) Initialize various VMD parameters, including the population size of the RSEA algorithm, the maximum number of iterations, and the upper and lower limits of the optimization parameters;

[0024] (22) The electric vehicle charging load data is introduced into the VMD model optimized by RSEA;

[0025] (23) When the current number of iterations is less than the maximum number of iterations, the algorithm starts iterating;

[0026] (24) The RSEA algorithm calculates the sample entropy and envelope entropy under different penalty factors alpha and decomposition levels k, and saves the solutions that meet the conditions as the Pareto optimal solution set;

[0027] The penalty factor alpha and the number of decomposition layers k are used as optimization parameters:

[0028]

[0029] In the formula, f VMD is the VMD model, y VMD is the decomposition result of VMD, alpha max and alpha min The upper and lower limits of the penalty factor alpha, K max and K min are the upper and lower limits of the decomposition level k respectively;

[0030] Taking sample entropy and envelope entropy as fitness function, sample entropy is expressed as:

[0031]

[0032] In the formula, SampEN is the sample entropy of each component, data is the time series data, m is the embedding dimension, r is the similarity tolerance, and B m (r) represents the probability that two vectors match m real numbers under similarity tolerance r; envelope entropy E p It is expressed as:

[0033]

[0034] Where Ep is the envelope entropy of each component, p j is the normalized form of a(j), where a(j) is the envelope signal obtained after Hilbert demodulation;

[0035] (25) The loop ends when the number of iterations reaches the maximum;

[0036] (26) The best set of solutions after normalization is selected, alpha and k are saved, and the algorithm ends.

[0037] Furthermore, the neural network model of the enhanced hidden layer in step (3) includes an input layer, an enhanced hidden layer and an output layer; the hidden layer is trained without considering retraining, the input weight of each temporary hidden layer is randomly generated, and multiple feature maps are input to allow the hidden layer to learn from several random linear maps; the enhanced hidden layer h:

[0038]

[0039] Among them, the elements {w, b} refer to the input weight matrix and bias vector of each feature map respectively; each temporary hidden layer is multiplied by a discount factor γ∈[0,1], and f is the activation function.

[0040] Furthermore, the implementation process of step (4) is as follows:

[0041] The number of neurons HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer are optimized. The optimization objectives are as follows:

[0042]

[0043] In the formula, f NAHL is the NAHL model, y NAHL The predicted results of NAHL under this parameter, HN, THN and P n They represent the number of neurons in the hidden layer HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer, and n is the number of layers; HN maxand HN min are the upper and lower limits of the number of neurons in the hidden layer, THN max and THN min are the upper and lower limits of the number of neurons in the temporary hidden layer, respectively;

[0044] The fitness function is set to use the root mean square error and symmetric mean absolute percentage error between the predicted and actual values ​​of the electric vehicle charging load as model evaluation indicators:

[0045]

[0046] Where Y i and i They represent the predicted value and actual value of electric vehicle charging load respectively; f1 is the root mean square error, and f2 is the symmetric mean absolute percentage error.

[0047] Furthermore, the implementation process of step (5) is as follows:

[0048] (51) Parameter initialization: Initialize the parameters of the RSEA algorithm, including the number of hidden layer neurons HN, the number of temporary hidden layer neurons THN, and the discount factor P of each temporary hidden layer;

[0049] (52) Population initialization: The population distribution is initialized using the Tent-Logistic-Cosine chaotic map to obtain the initial population PF, and the solution in the PF is projected into the two-dimensional radial space. The high-dimensional points are projected into the two-dimensional radial space through two projection weight vectors. The solution in the high-dimensional target space is first projected into the two-dimensional radial space. The two weight vectors are:

[0050] W1=(cos(θ1),cos(θ2),...,cos(θ m )) T

[0051] W2=(sin(θ1),sin(θ2),...,sin(θ m )) T

[0052] The coordinates of point A in radial space are expressed as:

[0053] Y i =(AW1(AI) -1 ,AW2(AI) -1 )

[0054] Where W1 and W2 represent weight vectors, A represents each individual in the population, and I represents an m×1 vector;

[0055] (53) Adaptive meshing in radial space:

[0056] Each area is considered as a small search space. These projected solutions are located in different rectangles through the adaptive grid partitioning method. Then, the better converged solutions in fewer rectangles are selected as parent individuals in the mating selection, and the local optimal solution in the solution space is searched more carefully, and finally the global optimal solution is led.

[0057] (54) Matching selection:

[0058] A binary tournament strategy is used to select individuals from PF, using the grid crowding in radial space and the convergence in target space as two indicators; the crowding degree of the rectangle is defined as Crowd(G s )=|S|,G s represents the selected grid, |S| represents the number of individuals in the grid;

[0059] (55) Generate offspring individuals:

[0060] The parent population obtained by matching selection simulates binomial crossover mutation and polynomial mutation, and adopts an exponentially decreasing adaptive mutation probability strategy to obtain the offspring population Q;

[0061] (56) Environmental selection:

[0062] The current population is sorted by the non-dominated sorting method to form a population P, and then the solutions in P are projected into the two-dimensional radial space, and the adaptive grid division in the radial space is performed to calculate the coordinates in the radial space and their positions in the segmented grid; then the m extreme value solutions of the population P are selected for the population Q, and the corresponding crowding degree is updated; according to their crowding degree and fitness value, the remaining solutions in P are selected one by one, and individuals with better fitness are retained; the fitness function is:

[0063] Fit(X,Q)=Con(X)·r P ·m-min||Y(x)-Y(Q)||

[0064]

[0065] Where Q is the set of selected solutions, Con(X) is the convergence of the solution X according to the equation, where Y(*) represents the coordinates of * in radial space, and the penalty parameter r P Adaptively adjust as the total evaluation progresses, t is the number of evaluated solutions, t max is the maximum number of evaluations; r P It becomes smaller as the degree of the true function increases;

[0066] (57) Determine whether the termination condition is met. If the number of iterations reaches the maximum, the loop ends; if not, continue iterating until the termination condition is met;

[0067] (58) Obtain the best solution to the problem and output various parameters and prediction results.

[0068] A device according to the present invention includes a memory and a processor, wherein:

[0069] A memory for storing computer programs that can be run on the processor;

[0070] A processor is used to execute the steps of the electric vehicle charging load prediction method based on NAHL as described above when running the computer program.

[0071] A storage medium described in the present invention stores a computer program, and when the computer program is executed by at least one processor, the steps of the electric vehicle charging load prediction method based on NAHL as described above are implemented.

[0072] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are:

[0073] 1. The present invention adopts a neural network with enhanced hidden layer (NAHL) to predict the charging load of electric vehicles, which can obtain more accurate prediction results; the characteristic analysis and prediction of the charging load of electric vehicles can better plan and manage the use of charging facilities and improve the utilization rate of charging facilities; it can better dispatch power resources and help users better arrange charging plans;

[0074] 2. The present invention uses the sample entropy and envelope entropy of the optimized VMD as the VMD decomposition model indicators, and the root mean square error (RMSE) and symmetric mean absolute percentage error (SMAPE) between the predicted value and the actual value of the electric vehicle charging load as the load forecasting model evaluation indicators; a set of solutions is found by considering multiple objectives instead of focusing on a single objective;

[0075] 3. The present invention proposes an improved RSEA algorithm, which uses the Tent-Logistic-Cosine chaotic mapping to initialize the population distribution, and proposes an exponentially decreasing adaptive mutation probability strategy; the sampling points generated by the mapping under the same conditions are evenly distributed in the entire distribution range, which can collect more information, avoid the problem of falling into the local optimality during the algorithm iteration process, and improve the diversity and convergence of the algorithm; compared with the commonly used chaotic mapping, it can improve the diversity of the population and the solution accuracy of the algorithm. Through comparison of experimental results, it is found that the improved RSEA algorithm has better performance and can effectively improve the accuracy of the electric vehicle charging load prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flow chart of the electric vehicle charging load prediction method based on NAHL;

[0077] Figure 2 Schematic diagram of the neural network structure with enhanced hidden layer;

[0078] Figure 3 is the Pareto front of a neural network with enhanced hidden layers;

[0079] Figure 4 The electric vehicle charging load prediction data output by the present invention is adopted. DETAILED DESCRIPTION

[0080] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0081] like Figure 1 As shown, the present invention proposes a method for predicting electric vehicle charging load based on NAHL, which specifically includes the following steps:

[0082] Step 1: Data collection, collect the historical data of EV charging load and divide it into training dataset and test dataset.

[0083] Collect the historical data of electric vehicle charging load. The selected data is the data of the Dutch electric vehicle smart charging pile. It is divided into a training data set and a test data set. The load of 24 hours is used as the input level, and the load at the next moment of the 24th hour is used as the output set. 80% of the historical data is used as the training set, and 20% of the historical data is used as the test set. Determine the training matrix and test matrix of each component. For the null value in the data, use y k =(y k-1 +y k+1 ) / 2 for filling; where y k is the filling data of the missing time, y k-1 is the data at the same time in the previous cycle, y k+1 The data is the same time in the next cycle.

[0084] Step 2: Variational Mode Decomposition (VMD) is used to decompose the electric vehicle charging load dataset, and the parameters of VMD are optimized using an improved radial space division based evolutionary algorithm (RSEA).

[0085] The penalty factor alpha and the number of decomposition layers k are used as optimization targets:

[0086]

[0087] In the formula, f VMDis the VMD model, y VMD is the decomposition result of VMD under this parameter, alpha max and alpha min The upper and lower limits of the penalty factor alpha, K max and K min are the upper and lower limits of the decomposition level k, respectively, where alpha max =3000, alpha min =100, K max =10, K min =3.

[0088] Taking sample entropy and envelope entropy as fitness function, sample entropy is expressed as:

[0089]

[0090] In the formula, SampEN is the sample entropy of each component, data is the time series data, m is the embedding dimension, r is the similarity tolerance, and B m (r) represents the probability that two vectors match m real numbers with similarity tolerance r.

[0091] Envelope entropy E p It is expressed as:

[0092]

[0093] Where Ep is the envelope entropy of each component, p j is the normalized form of a(j), where a(j) is the envelope signal obtained after Hilbert demodulation.

[0094] The improved RSEA optimizes the parameters of VMD. The specific process of the algorithm is as follows:

[0095] 1) Initialize various VMD parameters, including the population size of the RSEA optimization algorithm, the maximum number of iterations, and the upper and lower limits of the optimization parameters.

[0096] 2) The electric vehicle charging load data is introduced into the VMD model optimized by the improved RSEA.

[0097] 3) When the current number of iterations is less than the maximum number of iterations, the algorithm starts iterating.

[0098] 4) The RSEA algorithm calculates the sample entropy and envelope entropy under different penalty factors alpha and decomposition levels k, and saves the solutions that meet the conditions as the Pareto optimal solution set.

[0099] 5) The loop ends when the number of iterations reaches the maximum.

[0100] 6) The best set of solutions after normalization is selected, alpha and k are saved, and the algorithm ends.

[0101] Using the optimized VMD parameters to perform VMD decomposition, K-1 intrinsic mode function components and a residual component are obtained, and each intrinsic mode function is solved through iterative optimization. In each iteration, VMD will find an IMF and remove it from the signal, thereby decomposing the complex signal into several intrinsic mode functions.

[0102] Step 3: Build Figure 2 The Automatic artificial Neural network with an Augmented Hidden Layer (NAHL) model shown in the figure includes an input layer, an enhanced hidden layer, and an output layer. The hidden layer is trained without considering retraining, and the input weights of each temporary hidden layer are randomly generated. Multiple feature maps are input, allowing the hidden layer to learn from several random linear maps.

[0103] The new multi-feature map that can be input by NAHL advantage allows the hidden layer to learn multiple representations from several random linear mappings. The target output at time t can be calculated as follows:

[0104] 1) Set the network structure parameters, including the number of neurons HN, the number of neurons in the temporary hidden layer THN and the discount factor of each temporary hidden layer; the training sample set is x = [x1, x1, ..., x n ], the actual output is Y=[Y1,Y1,...,Y n ], the expected output is y=[y1,y1,...,y n ].

[0105] 2) Randomly generate the input weights w of hidden neurons i and bias b i , and expand the bias matrix to match the dimension of H. Use To represent the output of a neuron.

[0106] 3) The model has an enhanced hidden layer, that is, the hidden neuron output matrix h:

[0107]

[0108] The elements {w, b} refer to the input weight matrix and bias vector of each feature map, respectively. The temporary hidden layer is multiplied by a discount factor γ∈[0,1], and f is the activation function.

[0109] 4) Once the hidden layer is determined, the output y will be determined according to y = hβ; y is the output value expected by the model.

[0110] 5) Use the least squares method to calculate the output weight, the optimal weight β * It can be expressed as: * =h + y.h + is the generalized inverse matrix of the output matrix h.

[0111] 6) Get the actual output Y=hβ of the training data and calculate the training accuracy of the regression case.

[0112] Step 4: The root mean square error and symmetric mean absolute percentage error between the predicted value and the actual value of the electric vehicle charging load are used as the fitness function; and the training matrix and test matrix of each component are respectively brought into the established NAHL model.

[0113] RSEA is an evolutionary algorithm based on radial space partitioning, which aims to effectively solve optimization problems. The algorithm combines the evolutionary process of evolutionary algorithms and the idea of ​​radial space partitioning to improve search efficiency and global optimization ability. In order to improve the prediction accuracy of the model, the number of neurons HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer are the main parameters for optimization. The optimization goal is expressed as follows:

[0114]

[0115] In the formula, f NAHL is the NAHL model, y NAHL The predicted results of NAHL under this parameter, HN, THN and P n They represent the number of neurons in the hidden layer HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer, and n is the number of different layers. max and HN min are the upper and lower limits of the number of neurons in the hidden layer, THN max and THN min are the upper and lower limits of the number of neurons in the temporary hidden layer, respectively, where HN max =THN max =2000,HN min =THN min =100,

[0116] The fitness function is set as the root mean square error (RMSE) and symmetric mean absolute percentage error (SMAPE) between the predicted value and the actual value of the electric vehicle charging load as model evaluation indicators. The mathematical model is expressed as follows:

[0117]

[0118] Where Y i and i They represent the predicted value and actual value of electric vehicle charging load respectively; f1 is the root mean square error (RMSE), and f2 is the symmetric mean absolute percentage error (SMAPE).

[0119] Step 5: Use the improved RSEA algorithm to optimize and train the NAHL model, select the compromise solution of the Pareto frontier, and output the optimal parameters of the NAHL neural network sub-model; specifically, the following steps are included:

[0120] (1) Parameter initialization: Initialize the parameters of the RSEA algorithm, including the number of neurons in the hidden layer HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer.

[0121] (2) Population initialization. First, in order to generate an initial population with a wide range, uniform distribution, and strong global search capability, the Tent-Logistic-Cosine chaotic map is used to initialize the population distribution to obtain the initial population PF. The Tent-Logistic-Cosine chaotic map initialization strategy is as follows:

[0122]

[0123] Among them, x(i) is the current state, x(i+1) is the next state, r is the mapping parameter, r∈[0,1]. Then through PF i =PF min +x i * (PF max -PF min ) to obtain the initial result; PF max is the upper limit of the i-th decision variable, PF min is the lower limit of the ith decision variable, x i is the value initialized by the Tent-Logistic-Cosine chaotic map, PF i Represents the initialization result of the i-th decision variable.

[0124] The solution in PF is projected into the two-dimensional radial space. The high-dimensional point is projected into the two-dimensional radial space through two projection weight vectors. The solution in the high-dimensional target space is first projected into the two-dimensional radial space. The two weight vectors are:

[0125] W1=(cos(θ1),cos(θ2),...,cos(θ m )) T

[0126] W2=(sin(θ1),sin(θ2),...,sin(θ m )) T

[0127] The coordinates of point A in radial space represent Y i =(AW1(AI) -1 ,AW2(AI) -1 )

[0128] Where W1 and W2 represent weight vectors, A represents each individual in the population, and I represents an m×1 vector.

[0129] (3) Adaptive meshing in radial space:

[0130] By dividing the search space into different radial regions, the search accuracy can be improved. Each region can be considered as a small range of search space. These projected solutions are located in different rectangles through the adaptive grid partitioning method, and then the better converged solutions in fewer rectangles are selected as parent individuals in mating selection, which helps to search the local optimal solution in the solution space more carefully and eventually lead to the global optimal solution.

[0131] (4) Matching selection:

[0132] A binary tournament strategy is used to select individuals from PF, using the grid crowding in radial space and the convergence in target space as two indicators. The crowding degree of the rectangle is defined as Crowd(G s )=|S|,G s represents the selected grid, and |S| represents the number of individuals in the grid.

[0133] (5) Generate offspring individuals:

[0134] The parent population obtained by matching selection is subjected to simulated binomial crossover mutation and polynomial mutation to obtain the offspring population Q. The mutation operation is improved, and an exponentially decreasing adaptive mutation probability strategy is proposed. In the early stage of iteration, the mutation rate is high, and a widely distributed solution can be generated, which can improve the diversity of the algorithm; in the middle of the iteration, the mutation probability will gradually decrease; in the late stage of iteration, the mutation probability is the smallest, and the solutions obtained are more dense, which can improve the convergence of the algorithm. The exponentially decreasing adaptive mutation probability strategy is as follows:

[0135] MP=MP min *(MP max -MP min )^[1 / (1+K*FE / FE max )]

[0136] Where: K is the parameter that controls the exponential decrease. The larger K is, the faster the decrease is. maxis the maximum mutation probability, MP min is the minimum mutation probability, FE max is the maximum number of iterations, and FE is the current number of iterations. In the early stage of iteration, the mutation rate is high, and widely distributed solutions can be generated, which can improve the diversity of the algorithm; in the middle of the iteration, the mutation probability will gradually decrease; in the late stage of iteration, the mutation probability is the smallest, and the solutions obtained are more dense, which can improve the convergence of the algorithm.

[0137] (6) Environmental selection:

[0138] The environment selection first considers the elite non-dominated sorting, and then adopts the radial grid partitioning strategy as the second criterion. When the number of non-dominated individuals exceeds the maximum population size, the fitness of each individual in the grid with the largest crowding is calculated. The larger the fitness, the less help the individual provides to the maintenance of population diversity and convergence. The individual with the largest fitness is deleted, and the crowding of the grid is updated at the same time.

[0139] First, the current population is sorted by the non-dominated sorting method to form a population P, and then the solutions in P are projected into the two-dimensional radial space, and through the third step of adaptive meshing in the radial space, their coordinates in the radial space and their positions in the segmented grid are calculated. Then, m extreme value solutions of population P are selected for population Q, and the corresponding crowding is updated, where m is the number of targets. Finally, the remaining solutions in P are selected one by one according to their crowding degree and their fitness value, and individuals with better fitness are retained. The fitness function is:

[0140] Fit(X,Q)=Con(X)·r P ·m-min||Y(x)-Y(Q)||

[0141]

[0142] Where Q is the set of selected solutions, Con(X) is the convergence of the solution X according to the equation, and Y(*) represents the coordinates of * in radial space. In this equation, the penalty parameter r P Adaptively adjust as the total evaluation progresses, t is the number of evaluated solutions, t max is the maximum number of evaluations. This adaptive penalty-based approach allows the algorithm to maintain convergence in the early stages and have good diversity when converging. These populations are evenly distributed in the radial space.

[0143] In the formula, r P It becomes smaller as the number of true functions increases. The convergence of the target space and the crowding of the radial space are considered comprehensively to maintain the convergence and diversity of the population. The calculated fitness is the root mean square error (RMSE) and symmetric mean absolute percentage error (SMAPE) between the predicted and actual values ​​of the electric vehicle charging load.

[0144] (7) Termination condition check: Determine whether the termination condition is met. If the number of iterations reaches the maximum, the loop ends. If not, continue iterating until the termination condition is met.

[0145] (8) Output the optimal solution: Obtain the best solution to the problem and output various parameters and prediction results.

[0146] Step 6: Use the trained NAHL model and its corresponding test matrix in step 2 to calculate the prediction value of each NAHL model; aggregate the prediction values ​​of each NAHL model and output the final prediction result.

[0147] The root mean square error (RMSE), symmetric mean absolute percentage error (SMAPE), correlation coefficient (R), and coefficient of determination (R 2 ) and mean absolute error (MAE) were used to evaluate the performance of the VMD-RSEA-NAHL model. RMSE, MAE, R, R 2 The expressions of and MAE are as follows:

[0148]

[0149]

[0150]

[0151]

[0152] in:

[0153]

[0154]

[0155] Among them, n is the number of data, i represents the i-th data, Y i ,y i are the true value and predicted value of the data respectively.

[0156] The present invention also proposes a device, including a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the electric vehicle charging load prediction method based on NAHL as described above when running the computer program.

[0157] The present invention also proposes a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the electric vehicle charging load prediction method based on NAHL as described above are implemented.

[0158] The charging load after VMD decomposition is predicted. The results after optimization of the stationary component are not much different. The optimization of this model is mainly focused on the residual component. The residual component is predicted using the NAHL model, and the parameters of the NAHL model are optimized using the improved RSEA algorithm. The Pareto frontier diagram is as follows: Figure 3 Aggregate the predicted values ​​of each component and output the final prediction result as shown in Figure 4 As shown. Figure 4 The prediction effect can be verified more intuitively by Figure 4 It can be seen that the prediction results obtained by the proposed electric vehicle charging load prediction method are very consistent with the actual results, which proves the effectiveness of this model.

[0159] In order to further verify the prediction performance of the present invention, it is comprehensively compared with four models: back propagation neural network (BP), extreme learning machine (ELM), long short-term memory (LSTM), and least squares support vector machine (LSSVM). It is also compared with the optimization results of the single-objective algorithm PSO and the improved multi-objective algorithm RSEA. Table 1 lists the five statistical indicators of the proposed model and the comparison model, including RMSE, MAE, R, R 2 and the value of MAE.

[0160] Table 1 Error indicators of different models under the test set data sequence

[0161]

[0162]

[0163] Four statistical indicators of the comparison model, including RMSE, MAE, R, R 2 By comparing the five statistical indicators of different models, it is shown that the improved RSEA algorithm can effectively optimize the parameter selection of the NAHL model and thus improve the prediction accuracy.

[0164] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims

1. A NAHL-based electric vehicle charging load prediction method, characterized in that: The steps include: (1) Preprocessing the historical charging load data of electric vehicles and dividing it into a training data set and a test data set; (2) VMD is used to decompose and preprocess the charging load data of electric vehicles. The improved evolutionary algorithm based on radial space partitioning, namely, improved RSEA, is used to optimize the parameters of VMD, obtain K components and construct the corresponding K input matrices, and determine the training matrix and test matrix of each component; (3) Construct a neural network NAHL model with enhanced hidden layers; (4) The root mean square error and symmetric mean absolute percentage error between the predicted value and the actual value of the electric vehicle charging load are used as the fitness function; And the training matrix and test matrix of each component are respectively brought into the established NAHL model; (5) The improved RSEA algorithm is used to optimize and train the NAHL model, select the compromise solution of the Pareto frontier, and output the optimal parameters of the NAHL neural network sub-model; (6) Calculate the prediction value of each NAHL model using the trained NAHL model and its corresponding test matrix in step (2); (7) Aggregate the predicted values ​​of each NAHL model and output the final prediction result; The improved evolutionary algorithm based on radial space partitioning in step (2) is: Use Tent-Logistic-Cosine chaotic map initialization, the initialization strategy is as follows: Among them, x(i) is the current state, x(i+1) is the next state, r is the mapping parameter, r∈[0,1]; then through PF i =PF min +x i * (PF max -PF min ) to obtain the initial result; PF max is the upper limit of the i-th decision variable, PF min is the lower limit of the ith decision variable, x i is the value initialized by the Tent-Logistic-Cosine chaotic map, PF i represents the initialization result of the i-th decision variable; the mutation operation is improved, and an exponentially decreasing adaptive mutation probability strategy is proposed. The exponentially decreasing adaptive mutation probability strategy is as follows: MP=MP min *(MP max -MP min )^[1 / (1+K*FE / FE max )] In the formula, K is the parameter that controls the exponential decrease. The larger K is, the faster the decrease speed is. max is the maximum mutation probability, MP min is the minimum mutation probability, FE max is the maximum number of iterations, FE is the current number of iterations; in the early stage of iteration, the mutation rate is high, generating widely distributed solutions, improving the diversity of the algorithm; in the middle stage of iteration, the mutation probability will gradually decrease; in the late stage of iteration, the mutation probability is the smallest, the solutions obtained are more dense, improving the convergence of the algorithm; The neural network model of the enhanced hidden layer in step (3) includes an input layer, an enhanced hidden layer and an output layer; the hidden layer is trained without considering retraining, the input weight of each temporary hidden layer is randomly generated, and multiple feature maps are input to allow the hidden layer to learn from several random linear maps; the enhanced hidden layer h: Among them, the elements {w, b} refer to the input weight matrix and bias vector of each feature map respectively; each temporary hidden layer is multiplied by a discount factor γ∈[0,1], and f is the activation function.

2. The NAHL-based electric vehicle charging load prediction method according to claim 1, characterized in that: The preprocessing in step (1) is to fill in the empty values ​​in the data: and k (and k-1 +y k+1 ) / 2 Among them, y k is the filling data of the missing time, y k-1 is the data at the same time in the previous cycle, y k+1 The data is at the same time in the next cycle.

3. The NAHL-based electric vehicle charging load prediction method according to claim 1, characterized in that: The implementation process of step (2) is as follows: (21) Initialize various VMD parameters, including the population size of the RSEA algorithm, the maximum number of iterations, and the upper and lower limits of the optimization parameters; (22) The electric vehicle charging load data is introduced into the VMD model optimized by RSEA; (23) When the current number of iterations is less than the maximum number of iterations, the algorithm starts iterating; (24) The RSEA algorithm calculates the sample entropy and envelope entropy under different penalty factors alpha and decomposition levels k, and saves the solutions that meet the conditions as the Pareto optimal solution set; The penalty factor alpha and the number of decomposition layers k are used as optimization parameters: In the formula, f VMD is the VMD model, y VMD is the decomposition result of VMD, alpha max and alpha min The upper and lower limits of the penalty factor alpha, K max and K min are the upper and lower limits of the decomposition level k respectively; Taking sample entropy and envelope entropy as fitness function, sample entropy is expressed as: In the formula, SampEN is the sample entropy of each component, data is the time series data, m is the embedding dimension, r is the similarity tolerance, and B m (r) represents the probability that two vectors match m real numbers under similarity tolerance r; envelope entropy E p It is expressed as: Where Ep is the envelope entropy of each component, p j is the normalized form of a(j), where a(j) is the envelope signal obtained after Hilbert demodulation; (25) The loop ends when the number of iterations reaches the maximum; (26) The best set of solutions after normalization is selected, alpha and k are saved, and the algorithm ends.

4. The NAHL-based electric vehicle charging load prediction method according to claim 1, characterized in that: The implementation process of step (4) is as follows: The number of neurons HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer are optimized. The optimization objectives are as follows: In the formula, f NAHL is the NAHL model, y NAHL The prediction results of NAHL model, HN, THN and P n They represent the number of neurons in the hidden layer HN, the number of neurons in the temporary hidden layer THN, and the discount factor P of each temporary hidden layer. n For different layers; HN max and HN min are the upper and lower limits of the number of neurons in the hidden layer, THN max and THN min are the upper and lower limits of the number of neurons in the temporary hidden layer, respectively; The fitness function is set to use the root mean square error and symmetric mean absolute percentage error between the predicted and actual values ​​of the electric vehicle charging load as model evaluation indicators: Where Y if and if They represent the predicted value and actual value of electric vehicle charging load respectively; f1 is the root mean square error, and f2 is the symmetric mean absolute percentage error.

5. The NAHL-based electric vehicle charging load prediction method according to claim 1, characterized in that: The implementation process of step (5) is as follows: (51) Parameter initialization: Initialize the parameters of the RSEA algorithm, including the number of hidden layer neurons HN, the number of temporary hidden layer neurons THN, and the discount factor P of each temporary hidden layer; (52) Population initialization: The population distribution is initialized using the Tent-Logistic-Cosine chaotic map to obtain the initial population PF, and the solution in the PF is projected into the two-dimensional radial space. The high-dimensional points are projected into the two-dimensional radial space through two projection weight vectors. The solution in the high-dimensional target space is first projected into the two-dimensional radial space. The two weight vectors are: W1=(cos(θ1),cos(θ2),...,cos(θ m )) T W2=(sin(θ1),sin(θ2),...,sin(θ m )) T The coordinates of point A in radial space are expressed as: Y i =(AW1(AI) -1 ,AW2(AI) -1 ) Where W1 and W2 represent weight vectors, A represents each individual in the population, and I represents an m×1 vector; (53) Adaptive meshing in radial space: Each area is considered as a small search space. These projected solutions are located in different rectangles through the adaptive grid partitioning method. Then, the better converged solutions in fewer rectangles are selected as parent individuals in the mating selection, and the local optimal solution in the solution space is searched more carefully, and finally the global optimal solution is led. (54) Matching selection: A binary tournament strategy is used to select individuals from PF, using the grid crowding in radial space and the convergence in target space as two indicators; the crowding degree of the rectangle is defined as Crowd(G s )=|S|,G s represents the selected grid, |S| represents the number of individuals in the grid; (55) Generate offspring individuals: The parent population obtained by matching selection simulates binomial crossover mutation and polynomial mutation, and adopts an exponentially decreasing adaptive mutation probability strategy to obtain the offspring population Q; (56) Environmental selection: The current population is sorted by the non-dominated sorting method to form a population P, and then the solutions in P are projected into the two-dimensional radial space, and the adaptive grid division in the radial space is performed to calculate the coordinates in the radial space and their positions in the segmented grid; then the m extreme value solutions of the population P are selected for the population Q, and the corresponding crowding degree is updated; according to their crowding degree and fitness value, the remaining solutions in P are selected one by one, and individuals with better fitness are retained; the fitness function is: Fit(X,Q)=Con(X)·r P ·m-min||Y(X)-Y(Q)|| Where Q is the set of selected solutions, Con(X) is the convergence of the solution X according to the equation, Y(*) represents the coordinate of * in radial space, and the penalty parameter r P Adaptively adjust as the total evaluation progresses, t is the number of evaluated solutions, t max is the maximum number of evaluations; (57) Determine whether the termination condition is met. If the number of iterations reaches the maximum, the loop ends; if not, continue iterating until the termination condition is met; (58) Obtain the best solution to the problem and output various parameters and prediction results.

6. The NAHL-based electric vehicle charging load prediction method according to claim 5, characterized in that: The penalty parameter r P It becomes smaller as the degree of the true function increases.

7. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of the NAHL-based electric vehicle charging load prediction method as described in any one of claims 1 to 6 when running the computer program.

8. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by at least one processor, the steps of the NAHL-based electric vehicle charging load prediction method as described in any one of claims 1 to 6 are implemented.

Images