A method and system for predicting charging load of electric vehicles in a distribution network

Abstract

The present invention discloses a method for predicting the charging load of electric vehicles in a distribution network, and belongs to the technical field of urban power grids. The method comprises: calculating and preprocessing the charging load data of the temporal and spatial distribution based on the coordinate set of charging stations in the region, the historical charging load of the charging stations, and the vehicle location data in combination with the Thiessen polygon method; adaptively decomposing the preprocessed charging load data using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the rate of change of energy entropy, and obtaining the charging load data of the historical charging stations after decomposition; constructing a CycleNet model based on the charging load data of the historical charging stations; inputting the actual charging load data into the constructed CycleNet model to predict the charging load of the charging stations, thereby realizing the prediction of the charging load of electric vehicles in the distribution network. The present invention realizes the refined characterization of the temporal and spatial coupling characteristics of the load, improves the accuracy of the component decomposition of the non-stationary charging load signal, and improves the accuracy of the charging load prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban power grids, and more specifically, relates to a method and system for predicting charging load of electric vehicles in a distribution network. Background Art

[0002] With the rapid development of electric vehicles, the spatiotemporal distribution characteristics of charging loads are increasingly impacting distribution network operations. Traditional charging load forecasting methods lack geographic visualization and often treat spatial and temporal characteristics in isolation, resulting in inadequate capture of spatiotemporal interactions. These methods also fail to fully account for spatiotemporal coupling factors such as the geographic location of charging stations and the dynamic distribution of vehicles. This limits forecast accuracy and the lack of geographic visualization of spatiotemporal coupling, hindering subsequent geographic visualization for charging station site selection and service area optimization.

[0003] In the existing technology, charging load data often has noise interference, non-stationarity and abnormal fluctuations. The traditional empirical mode decomposition algorithm uses a fixed threshold for modal decomposition, which makes it difficult to adaptively adjust the envelope fitting parameters. As a result, the physical meaning of the decomposition components is unclear, making it difficult to effectively extract the intrinsic characteristics of the load.

[0004] In addition, conventional forecasting models (such as LSTM and ARIMA) lack the ability to collaboratively model periodic characteristics and nonlinear residual patterns, and cannot adapt to the complex dynamic changes of charging load. Summary of the Invention

[0005] To address the deficiencies in the prior art, the present invention provides a method and system for predicting the charging load of electric vehicles in a distribution network. Based on the coordinate set of charging stations within a region, the historical charging load data of the charging stations, and the vehicle location data within the region, the system combines the Thiessen polygon method to calculate the temporally and spatially distributed charging load data. The system then adaptively decomposes the preprocessed charging load data using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the rate of change of energy entropy to obtain the charging load data of historical charging stations. A CycleNet model is constructed based on the charging load data of historical charging stations. Actual charging load data is input into the constructed CycleNet model to predict the charging load of the charging stations, thereby realizing charging load prediction for electric vehicles in the distribution network. This method achieves a refined characterization of the temporal and spatial coupling characteristics of the load, improving the accuracy of the decomposition of the components of the non-stationary charging load signal. The constructed CycleNet model addresses the drawback of traditional models that insufficiently model the periodic component and residual component of the charging load of electric vehicle charging stations, significantly improving prediction accuracy.

[0006] The present invention adopts the following technical solutions.

[0007] A first aspect of the present invention provides a method for predicting charging load of electric vehicles in a distribution network, specifically comprising:

[0008] Based on the coordinate set of charging stations in the area, the historical charging load data of the charging stations, and the vehicle location data in the area, combined with the Thiessen polygon method, the charging load data with temporal and spatial distribution is calculated;

[0009] Pre-processing of charging loads calculated and distributed in time and space;

[0010] Adopting the adaptive empirical mode decomposition algorithm based on the dynamic average envelope and energy entropy change rate to adaptively decompose the pre-processed charging load data, the charging load data of the historical charging stations is obtained after decomposition;

[0011] Build a CycleNet model based on historical charging station charging load data;

[0012] The actual charging load data is input into the constructed CycleNet model to predict the charging load of charging stations and realize the charging load prediction of electric vehicles in the distribution network.

[0013] Preferably, the calculation of the temporal and spatial distribution of charging load data based on the charging station coordinate set in the area, the historical charging load data of the charging stations, and the vehicle position data in the area in combination with the Thiessen polygon method specifically includes:

[0014] Obtain the coordinate set of charging stations in the area, historical charging load data of charging stations, and vehicle location data in the area;

[0015] Solve the Thiessen polygons corresponding to all charging stations in the charging station coordinate set within the area;

[0016] According to the vehicle position data in the area and the solved Thiessen polygons, the number of vehicles in each Thiessen polygon is obtained;

[0017] Calculate the average charging power of each vehicle based on the historical charging load data of the charging station;

[0018] The benchmark load of the charging station is solved based on the number of vehicles in each Thiessen polygon and the average charging power of each vehicle. The benchmark load of the charging station is multiplied by the time weight to obtain the charging load data with both temporal and spatial distribution.

[0019] Preferably, the Thiessen polygons corresponding to all charging stations in the charging station coordinate set within the solution area specifically include:

[0020] Find all charging stations s in the area i The set of points whose distance to other charging stations is less than or equal to the distance to other charging stations is obtained, and the Thiessen polygons corresponding to all charging stations are obtained.

[0021] Preferably, the method of solving the benchmark load of the charging station according to the number of vehicles in each Thiessen polygon and the average charging power of each vehicle, multiplying the benchmark load of the charging station by the time weight, and solving the charging load data taking into account the temporal and spatial distribution includes:

[0022] According to the charging station s i The number of vehicles within the Thiessen polygon N i and the average charging power of the bicycle at time t r(t) Multiply them to get charging stations s i Base load at time t Q i (t) ;

[0023] The time weight is constructed based on the ratio of the historical average load at time t to the average load during the entire period;

[0024] The charging station s i Base load at time t Q i (t) Multiplying it with the time weight, we get the spatiotemporal distribution of charging load.

[0025] Preferably, the pre-processing of the charging load calculated based on the temporal and spatial distribution specifically includes:

[0026] The moving average method is used to zero-mean the charging load calculated and distributed in time and space to obtain zero-mean data;

[0027] Replace outliers in zero-mean data with the means of adjacent periods;

[0028] The replaced data is noise-smoothed and normalized to obtain preprocessed charging load data.

[0029] Preferably, the adaptive decomposition of the pre-processed charging load data using an adaptive empirical mode decomposition algorithm based on a dynamic average envelope and an energy entropy change rate specifically includes:

[0030] Initialize the residual signal based on the pre-processed charging load data, detect the extreme value of the current residual signal and correct the extreme value;

[0031] The charging pile online rate is set as the coefficient of the corrected extreme value, and the corrected extreme value is fitted with a B-spline envelope to solve the dynamic average envelope;

[0032] The difference between the current residual signal and the dynamic average envelope is set as the candidate component, and the normalized root mean square error and energy entropy change rate are calculated;

[0033] The standard deviation threshold and the energy entropy change rate are both less than the corresponding set thresholds as the stopping criterion. If the candidate component meets the stopping criterion, the current candidate component is set as the nth IMF component. Otherwise, the dynamic envelope fitting is repeated to update the candidate component and repeat the steps until the stopping criterion is met.

[0034] Determine whether the residual signal satisfies the monotonic function. If not, return the regenerated residual signal and perform cyclic decomposition of the residual signal until the residual signal satisfies the monotonic function. Then terminate the decomposition and obtain the decomposed historical charging load data set.

[0035] Preferably, the detecting and correcting the extreme value of the current residual signal specifically includes:

[0036] Detect the extreme value of the current residual signal. If the maximum value exceeds the rated power of the charging pile, correct the maximum value to the upper limit of the device. If the minimum value is negative, set the minimum value to zero.

[0037] Preferably, the CycleNet model is constructed based on the charging load data of historical charging stations, specifically including:

[0038] Normalize the input historical charging station charging load data and set the historical window length L in , the expected number of future time steps L out ;

[0039] Set up a learnable periodic function , where C is initialized to all zeros and L represents the cycle length;

[0040] According to the length L out The full 1 vector cyclic expansion of the periodic function can be used to learn the periodic component C repeat ;

[0041] Combined with the periodic component C repeat ,MLP is used to perform residual prediction on the charging load data of historical charging stations;

[0042] According to the residual prediction results and the period component C repeat Perform prediction synthesis to obtain charging load prediction results;

[0043] The mean square error is solved based on the predicted results of the charging load and the actual charging load value, and the mean square error is set as the loss function. The learnable cycle function C and the residual network parameters θ in the residual prediction process are trained with the optimizer to solve the optimal parameters and construct the CycleNet model.

[0044] Preferably, the combined periodic component C repeat , the MLP is used to perform residual prediction on the charging load data of historical charging stations, specifically including:

[0045] Charging load data x of historical charging stations history Normalize and intercept x history The last L in in Data x history [-L in :], where L in For the history window length, set L in is an integer multiple of L;

[0046] Calculate x history The last L in in Data x history [-L in :] and the periodic component C repeat The residual of

[0047] According to the calculated residual results, a two-layer MLP residual network is constructed to obtain the residual prediction results.

[0048] A second aspect of the present invention provides a distribution network electric vehicle charging load prediction system, which runs a distribution network electric vehicle charging load prediction method, including:

[0049] The charging load acquisition module is used to calculate the temporal and spatial distribution of charging load data based on the charging station coordinate set, historical charging load data of the charging stations, and vehicle location data in the area, combined with the Thiessen polygon method;

[0050] A pre-processing module, used to pre-process the charging load calculated and distributed in time and space;

[0051] An adaptive decomposition module is used to adaptively decompose the pre-processed charging load data using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the rate of change of energy entropy, and obtain the charging load data of the historical charging station after decomposition;

[0052] The model building module is used to build the CycleNet model based on the charging load data of historical charging stations;

[0053] The load forecasting module is used to input the actual charging load data into the constructed CycleNet model to predict the charging load of the charging station and realize the charging load prediction of electric vehicles in the distribution network.

[0054] Compared with the prior art, the beneficial effects of the present invention include at least:

[0055] By combining the Thiessen polygon method with vehicle location data, the service range of charging stations is dynamically divided, accurately quantifying the impact of regional vehicle distribution on load. This solves the problem of coarse spatial distribution modeling in traditional methods. Combined with time weight correction, a refined representation of the spatiotemporal coupling characteristics of load is achieved. The high degree of visualization of the Thiessen polygon method is conducive to subsequent geographic visualization optimization such as charging station planning and site selection, service range optimization, etc.

[0056] Adopting dynamic average envelope fitting and energy entropy change rate criterion based on the online rate of charging piles, the extreme value correction and envelope generation process of empirical mode decomposition are adaptively adjusted to effectively suppress modal aliasing and improve the decomposition accuracy of the components of non-stationary charging load signals.

[0057] The constructed CycleNet model explicitly models the periodic characteristics of the load through a learnable periodic function, and combines it with the MLP residual network to capture the nonlinear residual pattern. It solves the defect of traditional models in jointly modeling the periodic component and residual component of the charging load of electric vehicle charging stations, and significantly improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of a distribution network electric vehicle charging load prediction process provided in accordance with an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of constructing a spatiotemporal characteristic charging load process according to the Thiessen polygon method provided in an embodiment of the present invention;

[0060] Figure 3 3 is a schematic diagram of the adaptive empirical mode decomposition algorithm process provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0062] like Figure 1 As shown, embodiment 1 of the present invention provides a method for predicting charging load of electric vehicles in a distribution network, specifically comprising:

[0063] Step 1, such as Figure 2 As shown, based on the charging station coordinate set in the area, the historical charging load data of the charging stations, and the vehicle location data in the area, the charging load data with temporal and spatial distribution is calculated in combination with the Thiessen polygon method.

[0064] In a preferred but non-limiting embodiment of the present invention, step 1 comprises:

[0065] Step 1.1, solve the Thiessen polygons corresponding to all charging stations in the charging station coordinate set within the area;

[0066] Further preferably, step 1.1 includes:

[0067] (1)

[0068] Where,

[0069] V i Indicates all charging stations in the area i The set of points whose distance is less than or equal to the distance to other charging stations,

[0070] p represents a point in the region, Indicates that p is any point in a two-dimensional plane (region), represents a two-dimensional Euclidean space,

[0071] d(p,s i ) represents the point to charging station s in the area i distance,

[0072] d(p,s j ) represents the points in the area to the remaining charging stations s j distance.

[0073] Step 1.2: Superimpose the vehicle position data in the area with the Thiessen polygons calculated in step 1.1 to obtain the number of vehicles in each Thiessen polygon;

[0074] Step 1.3: Calculate the average charging power of a single vehicle based on the historical charging load data of the charging station. The average charging power of a single vehicle is expressed as follows:

[0075] (2)

[0076] Where,

[0077] r(t) represents the average charging power of a bicycle at time t,

[0078] L i (t) Indicates charging station s i The load at time t,

[0079] Σ L i (t) represents the sum of the loads of all charging stations at time t,

[0080] N veh Indicates the number of all vehicles.

[0081] In step 1.4, the benchmark load of the charging station is calculated based on the number of vehicles within each Thiessen polygon obtained in step 1.2 and the average charging power per vehicle obtained in step 1.3. This is multiplied by the historical charging load data of the charging station to obtain the charging load data that takes into account both temporal and spatial distribution.

[0082] Further preferably, step 1.4 includes:

[0083] Step 1.4.1, according to the charging station s i The number of vehicles within the Thiessen polygon N i and the average charging power of the bicycle at time t r(t) Multiply to get charging station s i Base load at time t Q i (t) ,

[0084] Step 1.4.2: construct the time weight based on the ratio of the historical average load at time t to the average load during the entire period;

[0085] Step 1.4.3, connect the charging station s i Base load at time t Q i (t) Multiplied by the time weight, the charging load with time and space distribution is obtained.

[0086] Step 2: pre-process the charging load calculated and distributed in time and space in step 1.

[0087] In a preferred but non-limiting embodiment of the present invention, step 2 comprises:

[0088] Step 2.1, use the moving average method to zero-mean the charging load calculated and distributed in time and space to obtain zero-mean data;

[0089] Step 2.2: Replace the outliers in the zero-mean data obtained in step 2.1 with the means of the adjacent time periods;

[0090] In step 2.3, the replaced data is noise-smoothed and normalized to obtain the preprocessed charging load data.

[0091] Step 3, such as Figure 3As shown, the charging load data preprocessed in step 2 is adaptively decomposed using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the energy entropy change rate to obtain the charging load data of the historical charging stations.

[0092] In a preferred but non-limiting embodiment of the present invention, step 3 comprises:

[0093] Step 3.1: Initialize the residual signal based on the preprocessed charging load data, detect the extreme value of the current residual signal and perform extreme value correction.

[0094] Further preferably, step 3.1 includes:

[0095] Step 3.1.1: Initialize the residual signal based on the preprocessed charging load data, which can be expressed as follows:

[0096] (4)

[0097] Where,

[0098] r0(t) represents the initialized residual signal,

[0099] x prep (t) Indicates the pre-processed charging load data.

[0100] Step 3.1.2, detect the current residual signal r n If the maximum value exceeds the rated power of the charging pile, the maximum value is corrected to the upper limit of the device. If the minimum value is negative (for example, but not limited to, caused by digital noise), the minimum value is set to zero.

[0101] Step 3.2: Set the charging pile online rate to the coefficient of the corrected extreme value, and perform B-spline envelope fitting based on the extreme value obtained in step 3.1 to obtain the dynamic average envelope. P env (t) , expressed as follows:

[0102] (5)

[0103] Where,

[0104] α(t) represents the online rate of charging piles at time t,

[0105] U(t) represents the charging power peak envelope generated by connecting the maximum points, i.e., the upper envelope;

[0106] L(t) represents the charging power valley envelope generated by connecting the minimum points, that is, the lower envelope.

[0107] In step 3.3, the difference between the current residual signal and the dynamic average envelope is set as the new candidate component, and the normalized root mean square error and energy entropy change rate are solved.

[0108] Further preferably, step 3.3 includes:

[0109] In step 3.3.1, the difference between the current residual signal and the dynamic average envelope is set as the candidate component h(t), which is expressed as follows:

[0110] (6)

[0111] Where,

[0112] r n (t) represents the current residual signal.

[0113] Step 3.3.2: Calculate the normalized root mean square error and energy entropy change rate based on the candidate components in step 3.3.1. The normalized root mean square error is expressed as follows:

[0114] (7)

[0115] Where,

[0116] SD k represents the standardized root mean square error,

[0117] h k (t) represents the candidate component of the kth iteration.

[0118] The rate of change of energy entropy is expressed as follows:

[0119] (8)

[0120] Where,

[0121] represents the rate of change of energy entropy,

[0122] H k represents the energy entropy of the k-th iteration,

[0123] represents the energy entropy of the candidate component,

[0124] Indicates time The ratio of the local energy to the total energy of the candidate component,

[0125] Indicates the current candidate component at time The value of

[0126] T represents the total length of the time series.

[0127] Step 3.4: Set the standard deviation threshold and the energy entropy change rate to be less than the corresponding set threshold as the stopping criterion. If the candidate component meets the stopping criterion, the current candidate component is set as the nth IMF component IMF n (t), otherwise, re-perform dynamic envelope fitting, update candidate components, and repeat the steps until the stopping criterion is met;

[0128] Step 3.5, according to step 3.4, get the nth IMF component IMF n (t), update the residual signal, which is expressed as follows:

[0129] (9)

[0130] Determine whether the residual signal satisfies the monotonic function. If not, return the regenerated residual signal to step 3.12 for cyclic decomposition of the residual signal until the residual signal satisfies the monotonic function. Terminate the decomposition and obtain the charging load data of the historical charging station.

[0131] Step 4: Build the CycleNet model based on the charging load data of historical charging stations.

[0132] In a preferred but non-limiting embodiment of the present invention, step 4 comprises:

[0133] Step 4.1: Normalize the input historical charging station charging load data and set the historical window length L in , the expected number of future time steps L out .

[0134] Step 4.2, set up a learnable periodic function , initialized to all zeros, L represents the cycle length, set to a daily cycle of 24 hours, that is, L=24, represents a real matrix of length L, represents the set of real numbers.

[0135] Step 4.3, by cyclically extending the learnable periodic function of step 4.2, the generation and prediction window length L out Aligned periodic component C repeat , expressed as follows:

[0136] (10)

[0137] Where,

[0138] C repeat represents the periodic component,

[0139] 1 Lout Indicates length L out A vector of all 1s,

[0140] L out Represents the prediction window length, which is set to an integer multiple of L and is used to represent the number of future time steps that the model expects to predict.

[0141] Step 4.4, combined with the periodic component C of step 4.3 repeat ,MLP is used to perform residual prediction on the charging load data of historical charging stations to obtain the residual prediction results.

[0142] Further preferably, step 4.4 includes:

[0143] Step 4.4.1, the charging load data x of the historical charging station history After normalization, the charging load data x of the historical charging station is intercepted history The last L in in Data x history [-L in :], L in is the history window length, which is used to represent the number of past time steps of the input model. Set L in is an integer multiple of L, for example but not limited to, when L=24, L in =96;

[0144] Step 4.4.2, calculate the charging load data x of the historical charging station history The last L in in Data x history [-L in :] and the periodic component C repeat The residual is expressed as follows:

[0145] (11)

[0146] Where,

[0147] x res represents the residual,

[0148] x history [-L in :] represents the charging load data x of the historical charging station history The last L in in data,

[0149] C repeat [-L in :] represents the periodic component C repeat The last L in in data.

[0150] In step 4.4.3, based on the calculated residual results, a two-layer MLP residual network is constructed, which is expressed as follows:

[0151] (12)

[0152] Where,

[0153] f θ (x res ) represents the residual prediction result, represents the residual network parameters, ,

[0154] W2 represents the linear transformation matrix that maps the hidden layer to the prediction residual,

[0155] ReLU represents the activation function,

[0156] x res represents the residual,

[0157] W1 represents the linear transformation matrix that maps the residual to the hidden layer,

[0158] b1 represents the activation threshold of the hidden layer neurons,

[0159] b2 represents the baseline shift of the output layer.

[0160] Step 4.5, based on the residual prediction result of step 4.4 and the period component C repeat Perform prediction synthesis, which is expressed as the following formula:

[0161] (13)

[0162] Where,

[0163] y pred Indicates the predicted value of charging load.

[0164] Step 4.6, calculate the mean square error based on the predicted value of the charging load obtained in step 4.5 and the actual charging load value, set the mean square error as the loss function, and combine the optimizer AdamW to optimize the learnable period function C and the residual network parameters. Perform training, find the optimal parameters, and build the CycleNet model.

[0165] Step 5: Input the actual charging load data into the CycleNet model constructed in step 4 to predict the charging load of the charging station and realize the charging load prediction of electric vehicles in the distribution network.

[0166] Data from an electric bus charging station in a certain city, comprising N charging piles and M buses, was collected. Every 15 minutes, the charging load and location data for the electric buses were collected, totaling 96 data points per 24 hours. The charging load and location data for a particular day were randomly selected from the electric bus charging station and predicted using a BP neural network, a wavelet neural network, and the present invention. The MAPE prediction accuracies achieved using the BP neural network, the wavelet neural network, and the present invention were 4.87%, 1.97%, and 0.99%, respectively. The BP neural network's prediction results showed significant fluctuations in predicted values and a poor ability to capture peaks and valleys. The wavelet neural network also exhibited delays during periods of rapid change. The present invention significantly outperformed both the BP neural network and the wavelet neural network.

[0167] Embodiment 2 of the present invention provides a distribution network electric vehicle charging load prediction system, which runs the distribution network electric vehicle charging load prediction method described in embodiment 1, including:

[0168] The charging load acquisition module is used to calculate the temporal and spatial distribution of charging load data based on the charging station coordinate set in the area, the historical charging load data of the charging stations, and the vehicle location data in the area, combined with the Thiessen polygon method;

[0169] A pre-processing module, used to pre-process the charging load calculated and distributed in time and space;

[0170] An adaptive decomposition module is used to adaptively decompose the pre-processed charging load data using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the rate of change of energy entropy, and obtain the charging load data of the historical charging station after decomposition;

[0171] The model building module is used to build the CycleNet model based on the charging load data of historical charging stations;

[0172] The load forecasting module is used to input the actual charging load data into the constructed CycleNet model to predict the charging load of the charging station and realize the charging load prediction of electric vehicles in the distribution network.

[0173] Compared with the prior art, the beneficial effects of the present invention include at least:

[0174] By combining the Thiessen polygon method with vehicle location data, the service range of charging stations is dynamically divided, accurately quantifying the impact of regional vehicle distribution on load. This solves the problem of coarse spatial distribution modeling in traditional methods. Combined with time weight correction, a refined representation of the spatiotemporal coupling characteristics of load is achieved. The high degree of visualization of the Thiessen polygon method is conducive to subsequent geographic visualization optimization such as charging station planning and site selection, service range optimization, etc.

[0175] Adopting dynamic average envelope fitting and energy entropy change rate criterion based on the online rate of charging piles, the extreme value correction and envelope generation process of empirical mode decomposition are adaptively adjusted to effectively suppress modal aliasing and improve the decomposition accuracy of the components of non-stationary charging load signals.

[0176] The constructed CycleNet model explicitly models the periodic characteristics of the load through a learnable periodic function, and combines it with the MLP residual network to capture the nonlinear residual pattern. It solves the defect of traditional models in jointly modeling the periodic component and residual component of the charging load of electric vehicle charging stations, and significantly improves the prediction accuracy.

[0177] The present disclosure may be a system, method and / or computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for predicting charging load of electric vehicles in a distribution network, characterized by: Based on the coordinate set of charging stations in the area, the historical charging load data of the charging stations, and the vehicle location data in the area, combined with the Thiessen polygon method, the temporal and spatial distribution of charging load data is calculated, including: The Thiessen polygons corresponding to all charging stations in the charging station coordinate set in the solution area include all charging stations in the solution area. s i The distance between the charging station and the charging station is less than or equal to the distance to other charging stations, and the corresponding Thiessen polygons of all charging stations are obtained; the number of vehicles within each Thiessen polygon is obtained by superimposing the vehicle position data in the area with the solved Thiessen polygons; the average charging power of each vehicle is solved based on the historical charging load data of the charging station; the benchmark load of the charging station is solved based on the number of vehicles within each Thiessen polygon and the average charging power of each vehicle, and the benchmark load of the charging station is multiplied by the time weight to obtain the charging load data with time and space distribution. Pre-processing of charging loads calculated and distributed in time and space; An adaptive empirical mode decomposition algorithm based on the dynamic average envelope and energy entropy change rate is used to adaptively decompose the preprocessed charging load data. After decomposition, the charging load data of the historical charging stations is obtained. Specifically, the algorithm includes: initializing the residual signal based on the preprocessed charging load data, detecting the extreme value of the current residual signal, and correcting the extreme value; The charging pile online rate is set as the coefficient of the corrected extreme value, and the corrected extreme value is fitted with a B-spline envelope to solve the dynamic average envelope; The difference between the current residual signal and the dynamic average envelope is set as the candidate component, and the normalized root mean square error and energy entropy change rate are solved according to the candidate component; The standard deviation threshold and the energy entropy change rate are both less than the corresponding set thresholds as the stopping criterion. If the candidate component meets the stopping criterion, the current candidate component is set as the nth IMF component. Otherwise, the dynamic envelope fitting is repeated to update the candidate component and repeat the steps until the stopping criterion is met. Determine whether the residual signal satisfies the monotonic function. If not, return the regenerated residual signal and perform cyclic decomposition of the residual signal until the residual signal satisfies the monotonic function. Then terminate the decomposition and obtain the charging load data of the historical charging station. Build a CycleNet model based on historical charging station charging load data; The actual charging load data is input into the constructed CycleNet model to predict the charging load of charging stations and realize the charging load prediction of electric vehicles in the distribution network.

2. A method for predicting charging load of electric vehicles in a distribution network according to claim 1, characterized in that: The benchmark load of the charging station is solved based on the number of vehicles and the average charging power of each vehicle within each Thiessen polygon, and the benchmark load of the charging station is multiplied by the time weight to obtain the charging load data with time and space distribution, including: According to the charging station s i The number of vehicles within the Thiessen polygon N i and the average charging power of the bicycle at time t r(t) Multiply them to get charging stations s i Base load at time t Q i (t) ; The time weight is constructed based on the ratio of the historical average load at time t to the average load during the entire period; The charging station s i Base load at time t Q i (t) Multiplying it with the time weight, we get the spatiotemporal distribution of charging load.

3. The method for predicting charging load of electric vehicles in a distribution network according to claim 1, wherein: The pre-processing of the charging load calculated and distributed in time and space specifically includes: The moving average method is used to zero-mean the charging load calculated and distributed in time and space to obtain zero-mean data; Replace outliers in zero-mean data with the means of adjacent periods; The replaced data is noise-smoothed and normalized to obtain preprocessed charging load data.

4. A method for predicting charging load of electric vehicles in a distribution network according to claim 3, characterized in that: The detecting and correcting the extreme value of the current residual signal specifically includes: Detect the extreme value of the current residual signal. If the maximum value exceeds the rated power of the charging pile, correct the maximum value to the upper limit of the device. If the minimum value is negative, set the minimum value to zero.

5. The method for predicting charging load of electric vehicles in a distribution network according to claim 1, characterized in that: The CycleNet model is constructed based on the charging load data of historical charging stations, specifically including: Normalize the input historical charging station charging load data and set the historical window length L in , the expected number of future time steps L out ; Set up a learnable periodic function , where C is initialized to all zeros and L represents the cycle length; According to the length L out The full 1 vector cyclic expansion of the periodic function can be used to learn the periodic component C repeat ; Combined with the periodic component C repeat ,MLP is used to perform residual prediction on the charging load data of historical charging stations; According to the residual prediction results and the period component C repeat Perform prediction synthesis to obtain charging load prediction results; The mean square error is solved based on the predicted results of the charging load and the actual charging load value, and the mean square error is set as the loss function. The learnable cycle function C and the residual network parameters θ in the residual prediction process are trained with the optimizer to solve the optimal parameters and construct the CycleNet model.

6. A method for predicting charging load of electric vehicles in a distribution network according to claim 5, characterized in that: The combined periodic component C repeat , the MLP is used to perform residual prediction on the charging load data of historical charging stations, specifically including: Charging load data x of historical charging stations history Normalize and intercept x history The last L in in Data x history [-L in :], where L in For the history window length, set L in is an integer multiple of L; Calculate x history The last L in in Data x history [-L in :] and the periodic component C repeat The residual of According to the calculated residual results, a two-layer MLP residual network is constructed to obtain the residual prediction results.

7. A distribution network electric vehicle charging load forecasting system, which runs a distribution network electric vehicle charging load forecasting method according to any one of claims 1 to 6, characterized in that: The charging load acquisition module is used to calculate the temporal and spatial distribution of charging load data based on the charging station coordinate set in the area, the historical charging load data of the charging stations, and the vehicle location data in the area, combined with the Thiessen polygon method; A pre-processing module, used to pre-process the charging load calculated and distributed in time and space; An adaptive decomposition module is used to adaptively decompose the pre-processed charging load data using an adaptive empirical mode decomposition algorithm based on the dynamic average envelope and the rate of change of energy entropy, and obtain the charging load data of the historical charging station after decomposition; The model building module is used to build the CycleNet model based on the charging load data of historical charging stations; The load forecasting module is used to input the actual charging load data into the constructed CycleNet model to predict the charging load of the charging station and realize the charging load prediction of electric vehicles in the distribution network.

Images