Short-term power load prediction method and system under new economic situation

Abstract

The invention discloses a short-term power load prediction method under a new economic situation, and the method comprises the steps: S1, determining a load prediction target region, collecting consumer behavior data, electric vehicle data, energy transformation related data, and digital intelligent data, and forming an original data set; s2, after data in the original data set is preprocessed, normalization processing is carried out, and a normalized data set is obtained; s3, performing data dimension reduction processing on the normalized data set, then constructing a VMD-GA model, optimizing parameters of a variational modal decomposition model by using a genetic algorithm, training the VMD-GA model by using the data set after dimension reduction processing, and obtaining a data set after modal decomposition; s4, constructing an IPSO-GRU model, optimizing hyper-parameters of the GRU model by using a particle swarm algorithm, and training the IPSO-GRU model by using the data set after modal decomposition; and S5, acquiring operation parameters of the target area, and performing power load prediction of the next target time period based on the VMD-GA model and the IPSO-GRU model. The design can adapt to the existing power utilization structure and load characteristics.

Description

Technical Field

[0001] The present invention relates to a short-term power load forecasting method and system under the new economic situation, which is specifically applicable to power load forecasting that adapts to the power consumption structure and load characteristics of a large number of distributed power sources, electric vehicles and intelligent power equipment. Background Art

[0002] Electricity has become one of the most important energy sources in today's society and plays an indispensable role in all fields. Short-Term Load Forecasting (STLF) refers to the prediction of electricity demand in the next few hours to days. It is one of the key tasks in the operation and management of power systems, especially for power companies, grid operators and energy dispatch centers.

[0003] At present, with the continuous development of the economy and the further adjustment of the industrial structure, the power consumption structure and load characteristics of my country's power grid have also changed accordingly. Under the new economic situation, the prediction of power load has become more complicated, mainly affected by various factors such as economic development, changes in industrial structure, energy transformation (such as wind power, solar energy), changes in consumer behavior, the popularization of electric vehicles, and digital intelligence (application of technologies such as smart home and Internet of Things). In order to better cope with these changes, short-term power load forecasting methods need to be continuously updated and optimized. There are many types of traditional short-term power load forecasting methods. For example, time series analysis models (such as ARIMA, SARIMA, etc.) are the earliest traditional statistical models used in short-term power load forecasting. Such models are simple and easy to implement, and are suitable for data with stable power load change trends and obvious seasonal laws. However, they cannot handle nonlinear and complex load changes well, and are sensitive to external influencing factors (such as sudden charging demand of electric vehicles or fluctuations in power grid load, etc.), and cannot effectively handle multivariate situations. In addition, regression analysis models (such as linear regression and multivariate regression) can incorporate multiple external factors into the model to improve the interpretability of the prediction, but they are currently only applicable to linear relationships and cannot handle complex nonlinear data.

[0004] In the load forecasting under the new economic situation, the power system is significantly affected by multiple uncertain factors, and the intricate interrelationships between these factors make it challenging to accurately predict load demand. As a new signal decomposition method, variational mode decomposition (VMD) can decompose the original load signal into multiple intrinsic mode functions (IMFs) and effectively extract different frequency components. When dealing with non-stationary and nonlinear time series, it will greatly improve the accuracy of load forecasting for complex nonlinear signals, thereby capturing power system load data under non-stationary and volatile conditions. On the other hand, genetic algorithms (GA) are based on the theory of biological evolution. By changing the search process through selection, crossover and mutation, they can explore and find the global optimal solution in multiple possible solution spaces, especially in complex power load forecasting models. It is used to optimize the weights and thresholds of the BP neural network to solve the problem of falling into the local optimal solution, while accelerating the training speed of the model, enhancing the adaptability and convergence of the model, and thus improving the recognition accuracy. Gated Recurrent Unit (GRU) is a variant of Recurrent Neural Network (RNN) commonly used to process sequence data. Compared with standard RNN, GRU can better capture power loads with long-term dependency characteristics such as periodicity and seasonality, and is particularly suitable for dealing with complex time series problems. At the same time, it can also improve the prediction accuracy by integrating multi-dimensional features, such as inputting external factors (such as consumer power consumption patterns, renewable energy generation fluctuations) and other multi-dimensional features. Compared with Long Short-Term Memory (LSTM), GRU has lower computational cost, faster training speed and model inference speed, and can achieve more efficient short-term power load forecasting. In addition, Improved Particle Swarm Optimization (IPSO) is a global optimization algorithm based on swarm intelligence that can find the optimal solution in multiple dimensions. IPSO is used to optimize the hyperparameters of the GRU network, such as learning rate, number of hidden layer nodes, etc. Through this optimization method, the model can automatically adjust parameters to improve prediction accuracy and generalization ability, avoiding the tedious work of manual parameter adjustment.

[0005] In summary, the present invention updates and optimizes the traditional load forecasting and provides a short-term load forecasting method under the new economic situation. That is, VMD-GA-IPSO-GRU is used for load forecasting, and the signal is decomposed and optimized by VMD-GA, combined with the hyperparameter optimization of IPSO and the modeling of time series data by GRU, so as to significantly improve the accuracy of power load forecasting, especially in the face of sudden load changes or complex nonlinear load trends under the new economic situation. The model will have stronger robustness and adaptability. Summary of the invention

[0006] The purpose of the present invention is to overcome the problem that the existing power load forecasting algorithm in the prior art cannot adapt to the new power consumption structure and load characteristics, and to provide a short-term power load forecasting method and system under the new economic situation that can adapt to the existing power consumption structure and load characteristics.

[0007] To achieve the above objectives, the technical solution of the present invention is:

[0008] In a first aspect, the present invention provides a method for short-term power load forecasting under a new economic situation, comprising the following steps:

[0009] S1. Determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a raw data set;

[0010] S2, after preprocessing and analyzing the data in the original data set, normalization is performed to obtain a normalized data set;

[0011] S3, performing data dimensionality reduction processing on the normalized data set, then constructing a VMD-GA model, using a genetic algorithm to optimize the parameters of the variational mode decomposition model, and using the data set after dimensionality reduction processing to train the VMD-GA model to obtain a data set after mode decomposition;

[0012] S4. Construct the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model. The IPSO-GRU model is trained using the data set after modal decomposition.

[0013] S5. Collect the operating parameters of the target area and make power load forecasts for the next target period based on the VMD-GA model and the IPSO-GRU model.

[0014] In said S1, a load forecast target area e is determined to form an original data set;

[0015] The original data set includes:

[0016] C Q =CQ 1 +C Q 2 +C Q 3 +C Q 4 +C Q 5

[0017] E Q =E Q 1 +E Q 2 +E Q 3 +E Q 4 +E Q 5

[0018] N Q =N Q 1 +N Q 2 +N Q 3 +N Q 4

[0019] I Q =I Q 1 +I Q 2 +I Q 3 +I Q 4

[0020] Where: C Q For consumer behavior data, C Q 1 is the industrial electricity consumption data, C Q 2 is household electricity consumption data, C Q 3 is the commercial electricity consumption data, C Q 4 The usage pattern of home devices on the cloud platform, C Q 5 is the user behavior history data, E Q For electric vehicle data, E Q 1 For electric vehicle charging pile data, E Q 2 Charging behavior data for electric vehicles, E Q 3 is the charging period data, E Q4 Charging capacity for electric vehicles, E Q 5 is the location data of the charging station, N Q For data related to energy transformation, N Q 1 is the wind power generation data, N Q 2 is the solar power generation data, N Q 3 is the energy policy related data, N Q 4 is the energy storage system data, I Q To digitize intelligent data, I Q 1 For smart meter data, I Q 2 For sensor and IoT data, I Q 3 Real-time power consumption data for IoT devices, Q 4 For cloud computing platform data.

[0021] In S2, the data in the original data set is preprocessed and analyzed, and missing data or abnormal data are supplemented according to the following formula:

[0022] C Q (d,t)=a c1 C Q (d-1,t)+a c2 C Q (d+1,t)+a c3 C Q (d,t-1)+a c4 C Q (d,t+1)

[0023] E Q (d,t)=a e1 E Q (d-1,t)+a e2 E Q (d+1,t)+a e3 E Q (d,t-1)+a e4 E Q (d,t+1)

[0024] N Q (d,t)=a n1 N Q (d-1,t)+a n2 N Q (d+1,t)+a n3 N Q(d,t-1)+a n4 N Q (d,t+1)

[0025] I Q (d,t)=a i1 I Q (d-1,t)+a i2 I Q (d+1,t)+a i3 I Q (d,t-1)+a i4 I Q (d,t+1)

[0026] Among them, C Q (d,t),E Q (d,t),N Q (d,t),I Q (d, t) represent the consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data at the t-th time on the d-th day that need to be corrected; a c1 、a c2 、a c3 、a c4 are the weight coefficients for completing consumer behavior data; a e1 、a e2 、a e3 、a e4 are weight coefficients for completing data related to energy transformation; a n1 、a n2 、a n3 、a n4 are the weight coefficients for electric vehicle data completion; a i1 、a i2 、a i3 、a i4 They are all weight coefficients for digital and intelligent data completion;

[0027] Normalize the preprocessed data to get the normalized data set:

[0028]

[0029] Among them, C Q* 、E Q* 、N Q* ,I Q* They are normalized consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data.

[0030] In S3, each data set C is first Q* , E Q* , N Q* , IQ* Perform principal component analysis to reduce dimension. For the data set X∈{C Q* , E Q* , N Q* , I Q* = R N×P , where N is the number of samples and P is the feature dimension. First, center the data X centered =X-μx, μx is the mean vector of each feature, and the covariance matrix is ​​calculated:

[0031]

[0032] Perform eigendecomposition on the covariance matrix:

[0033] ∑{C Q* , E Q* , N Q* , I Q*} i =∑Xν i =λ i ν i

[0034] Among them, v i represents the eigenvector, λ i represents the corresponding eigenvalue, X centered It is a centralized dataset;

[0035] Sort the eigenvectors from large to small according to the eigenvalues ​​of the covariance matrix, select the first h eigenvectors, and form the dimension reduction matrix W h ∈R P×h , and project the data onto the principal components; by Q* , E Q* , N Q* , I Q*} to reduce the dimension, we can get the reduced dimension data set X′:

[0036] X′=X centered W h ∈R N×h ∈{C Q* , E Q* , N Q* , I Q*}′

[0037] The dimension reduction data set X′ is subjected to VMD decomposition, that is, f(t,X′) is taken as input, which is the sum of the data of the distribution network in area e during period t, and is decomposed into m modal components u with specific sparsity m(t), and determine the center frequency and bandwidth of each modal component at the same time. The nonlinear time series characteristics of the data can be effectively extracted by the variational solution process that transforms the data from the time domain to the frequency domain. The variational solution process decomposes the original data into M modal components, minimizing the sum of the estimated bandwidths of each modal component. The constrained variational model is as follows:

[0038]

[0039] Where M is the number of modal components, z m (t) is the time domain function of the mth subsequence electric power component decomposed from the original electric power, w m is the inherent center frequency of the mth subsequence electric power component, δ(t) is the change rate of the electric power in the decomposed dimensionality reduction data set, j represents the normal complex value of the original electric power, represents the partial derivative with respect to time t;

[0040] In order to obtain the optimal solution of the above variational model, the following formula is used:

[0041]

[0042] Among them, u m (t) represents the time domain function of the mth mode obtained by decomposition, w m represents the center frequency of the mth mode, α is a balance parameter used to control the weight of the bandwidth constraint, Indicates z m (t) is the square of the norm, λ(t) represents the Lagrange multiplication, and <> represents the inner product operation;

[0043] To simplify the calculation, the Lagrangian function is converted to the frequency domain for processing, that is, u m The optimization of (t) is transferred from the time domain to the frequency domain to obtain The frequency domain function of the mth mode obtained by level decomposition; the corresponding z m The frequency domain corresponding to (t) is That is, the frequency domain function of the mth subsequence electric power component decomposed from the original electric power; the frequency domain multiplier corresponding to the λ(t) Lagrange multiplier is Then in the frequency domain:

[0044]

[0045] Then, the alternating multiplier direction algorithm is used to iterate continuously, and the optimal solution can be obtained as follows:

[0046]

[0047]

[0048] Among them, n is the number of iterations, and τ is the step size parameter for updating the Lagrange multiplier.

[0049] In S3, a genetic algorithm is used to optimize the parameters of the variational mode decomposition model. In the genetic algorithm, for C Q′ 、E Q′ 、N Q′ ,I Q′ The fitness of each data set can be expressed as follows:

[0050]

[0051] Among them, u Cm (t),u Em (t),u Nm (t),u Im (t) are C Q′ 、E Q′ 、N Q′ ,I Q′ The modes of the data set obtained by VMD decomposition; the optimal mode number after genetic algorithm optimization is as follows:

[0052]

[0053] Among them, M C ,M E ,M N ,M I represents the optimal number of modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) represents the best modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) together constitute the data set after modal decomposition.

[0054] In S4, the particle swarm algorithm is used to optimize the hyperparameters of the GRU model:

[0055] Set some hyperparameters of GRU, initialize IPSO parameters, find the optimal solution through IPSO iteration, input the optimized hyperparameters into GRU model training, and output the predicted value. Its mathematical description is shown in the following formula;

[0056] z t =σ(W z·[h t-1 ,x t ])

[0057] r t =σ(W r ·[h t-1 ,x t ])

[0058] h t =φ(W·[r t *h t-1 ,x t ])

[0059] h t =(1-z t )*h t-1 +z t *h t

[0060] Among them, x t is the input vector at time step t, using the data set C after modal decomposition Q (t), E Q (t), N Q (t), I Q (t) as the input vector, h t-1 and h t are the state variables of the previous moment and this moment respectively, r t is the update gate, z t is the reset gate, h t is a candidate set; W r , W z , W h is the weight parameter; σ is the Sigmoid activation function; φ is the Tanh activation function;

[0061] The mathematical descriptions of σ and φ are as follows:

[0062]

[0063] The particle swarm algorithm is used to optimize some hyperparameters of the GRU model. Q′ 、E Q′ 、N Q′ ,I Q′ Data set, output prediction value as follows:

[0064]

[0065] in, All are bias items, which fuse the data sets:

[0066]

[0067] Output of short-term power load forecasts

[0068] In a second aspect, the present invention provides a short-term power load forecasting system under the new economic situation, the system is used to execute the short-term power load forecasting method under the new economic situation, specifically comprising: a data acquisition module, a data preprocessing module, a VMD-GA model building module, an IPSO-GRU model building module, and a power load forecasting module;

[0069] The data collection module is used to determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a collection of original data;

[0070] The data preprocessing module is used to perform preprocessing and analysis on the data in the original data set, and then perform normalization processing to obtain a normalized data set;

[0071] The VMD-GA model building module is used to perform data dimensionality reduction processing on the normalized data set, then build the VMD-GA model, use the genetic algorithm to optimize the parameters of the variational mode decomposition model, and use the data set after dimensionality reduction processing to train the VMD-GA model to obtain the data set after mode decomposition;

[0072] The IPSO-GRU model building module is used to build the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model and train the IPSO-GRU model using the data set after modal decomposition.

[0073] The power load forecasting module is used to collect the operating parameters of the target area and forecast the power load for the next target period based on the VMD-GA model and the IPSO-GRU model.

[0074] In a third aspect, the present invention provides a short-term power load forecasting device under the new economic situation, comprising a memory and a processor, wherein the memory is used to store computer program code and transmit the computer program code to the processor;

[0075] The processor is used to execute the above-mentioned short-term power load forecasting method under the new economic situation according to the instructions in the computer program code.

[0076] In a fourth aspect, the present invention provides a computer program product, including a computer program, which is executed by a processor to implement the aforementioned short-term power load forecasting method under the new economic situation.

[0077] In a fifth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the aforementioned method for short-term power load forecasting under the new economic situation.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] 1. The present invention provides a short-term power load forecasting method under the new economic situation. The data sets such as consumer behavior, electric vehicle charging, energy transformation, and digital intelligent power grid are assigned and calculated, so that the forecasting model has a high responsiveness to the dynamic changes of such factors. The traditional load forecasting method is updated and optimized, which can effectively cope with the complexity and variability of power load forecasting under the new economic situation.

[0080] 2. The present invention incorporates multi-dimensional data sources into a short-term power load forecasting method under the new economic situation, so that the model can more comprehensively capture the changing trend of power load. At the same time, the refined data set and segmented power consumption pattern will enable the power grid data forecast under the new economic situation to have higher temporal and spatial accuracy to cope with load fluctuations in critical periods, thereby improving the reliability and stability of the model prediction, so as to adapt to the needs of the development of new smart grids, and have strong market competitiveness and broad application prospects.

[0081] 3. In the short-term power load forecasting method based on VMD-GA-IPSO-GRU under a new economic situation of the present invention, VMD-GA can improve the data decomposition effect and enhance the model's ability to cope with complex power load characteristics. IPSO-GRU can help the model better adapt to complex time series data, cope with the uncertainties brought by electric vehicles, energy transformation and digital intelligent power grids, and improve the stability of short-term load forecasting. The optimized model also has the ability to efficiently process large-scale data, and can quickly train and predict while ensuring accuracy, adapting to the needs of power grid intelligence for real-time prediction.

[0082] 4. A short-term power load forecasting system under a new economic situation of the present invention includes a data acquisition module, a data preprocessing module, a VMD-GA model building module, an IPSO-GRU model building module, and a power load forecasting module. The system is used to implement the steps of the short-term power load forecasting method under a new economic situation provided in any of the above technical solutions. Therefore, the system also includes all the beneficial effects of the short-term power load forecasting method under a new economic situation provided in any of the above technical solutions, which will not be repeated here.

[0083] 5. A short-term power load forecasting device under a new economic situation of the present invention includes a processor and a memory, the memory is used to store computer program code, and transmit the computer program code to the processor, and the processor is used to execute the short-term power load forecasting method under a new economic situation provided in any of the above technical solutions according to the instructions in the computer program code. Therefore, the device also includes all the beneficial effects of the short-term power load forecasting method under a new economic situation provided in any of the above technical solutions, which will not be repeated here.

[0084] 6. A computer program product of the present invention, when executed by a processor, implements the steps of the short-term power load forecasting method under the new economic situation provided in any of the above technical solutions. Therefore, the computer program product also includes all the beneficial effects of the short-term power load forecasting method under the new economic situation provided in any of the above technical solutions, which will not be repeated here.

[0085] 7. A computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the steps of the short-term power load forecasting method under the new economic situation provided in any of the above technical solutions. Therefore, the computer-readable storage medium also includes all the beneficial effects of the short-term power load forecasting method under the new economic situation provided in any of the above technical solutions, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flow chart of the method of the present invention.

[0087] Figure 2 It is a system diagram of the present invention.

[0088] Figure 3 It is a diagram of the equipment of the present invention.

[0089] Figure 4 It is a decomposition prediction diagram of the preprocessed data in Example 1.

[0090] Figure 5 This is a comparison diagram of the effects in Example 1. DETAILED DESCRIPTION

[0091] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0092] Embodiment 1:

[0093] See also Figure 1 , a short-term power load forecasting method under the new economic situation, comprising the following steps:

[0094] S1. Determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a raw data set;

[0095] Determine the load forecast target area e and form an original data set;

[0096] The original data set includes:

[0097] C Q =C Q 1 +C Q 2 +C Q 3 +C Q 4 +C Q 5

[0098] E Q =E Q 1 +E Q 2 +E Q 3 +E Q 4 +E Q 5

[0099] N Q =N Q 1 +N Q 2 +N Q 3 +N Q 4

[0100] I Q =I Q 1 +I Q 2 +I Q 3 +I Q 4

[0101] Where: C Q For consumer behavior data, C Q 1 is the industrial electricity consumption data, C Q 2 is household electricity consumption data, C Q 3 is the commercial electricity consumption data, C Q 4The usage pattern of home devices on the cloud platform, C Q 5 is the user behavior history data, E Q For electric vehicle data, E Q 1 For electric vehicle charging pile data, E Q 2 Charging behavior data for electric vehicles, E Q 3 is the charging period data, E Q 4 Charging capacity for electric vehicles, E Q 5 is the location data of the charging station, N Q For data related to energy transformation, N Q 1 is the wind power generation data, N Q 2 is the solar power generation data, N Q 3 is energy policy related data, N Q 4 is the energy storage system data, I Q To digitize intelligent data, I Q 1 For smart meter data, I Q 2 For sensor and IoT data, I Q 3 Real-time power consumption data for IoT devices, Q 4 For cloud computing platform data.

[0102] In S1, collect consumer behavior data C Q (Including industrial electricity data C Q 1 、Household electricity consumption data C Q 2 、Commercial electricity consumption data C Q 3 、Cloud platform home device usage mode C Q 4 , User behavior history data C Q 5 The key features are the electricity consumption time period, consumption pattern (such as peak time, off-peak time), electricity load fluctuation, holidays, and the impact of climate conditions on consumer behavior; collect electric vehicle data E Q (Including electric vehicle charging pile data E Q 1 , Electric vehicle charging behavior E Q2 , Charging period data E Q 3 、Charge E Q 4 、Charging station location E Q 5 etc.), pay attention to the charging time and amount of electric vehicles, the distribution and usage patterns of electric vehicles (such as the difference between charging during the day and at night), and the impact of electric vehicle charging on load; collect data related to energy transformation Q (including renewable energy such as wind power Q 1 、Solar power generation data N Q 2 、Energy policy related data Q 3 , Energy storage system data N Q 4 ), the key features include renewable energy generation (large fluctuations), load regulation of power storage systems, seasonal changes in power load (such as load differences between summer and winter); collecting digital intelligent data I Q (Including smart grid devices such as smart meter data Q 1 , Sensors and IoT Data I Q 2 , Real-time power consumption data of IoT devices Q 3 、Cloud Computing Platform DataI Q 4 The key features are real-time load data, dynamic grid adjustment, demand response data, and load patterns of smart appliances.

[0103] S2, after preprocessing and analyzing the data in the original data set, normalization is performed to obtain a normalized data set;

[0104] In S2, the data in the original data set is preprocessed and analyzed, and missing data or abnormal data are supplemented according to the following formula:

[0105] C Q (d,t)=a c1 C Q (d-1,t)+a c2 C Q (d+1,t)+a c3 C Q (d,t-1)+a c4 C Q (d,t+1)

[0106] E Q (d,t)=ae1 E Q (d-1,t)+a e2 E Q (d+1,t)+a e3 E Q (d,t-1)+a e4 E Q (d,t+1)

[0107] N Q (d,t)=a n1 N Q (d-1,t)+a n2 N Q (d+1,t)+a n3 N Q (d,t-1)+a n4 N Q (d,t+1)

[0108] I Q (d,t)=a i1 I Q (d-1,t)+a i2 I Q (d+1,t)+a i3 I Q (d,t-1)+a i4 I Q (d,t+1)

[0109] Among them, C Q (d,t),E Q (d,t),N Q (d,t),I Q (d, t) represent the consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data at the t-th time on the d-th day that need to be corrected; a c1 、a c2 、a c3 、a c4 are the weight coefficients for completing consumer behavior data; a e1 、a e2 、a e3 、a e4 are weight coefficients for completing data related to energy transformation; a n1 、a n2 、a n3 、a n4 are the weight coefficients for electric vehicle data completion; a i1 、a i2 、a i3 、a i4 They are all weight coefficients for digital and intelligent data completion;

[0110] Normalize the preprocessed data:

[0111] Electric vehicle data, energy transformation-related data, and digital intelligent data, a is the weight coefficient. The farther from the target point, the smaller a is. In addition, abnormal data is corrected and data is normalized. The category data can be converted into data format using One-Hot encoding or label encoding, and the time series data is Min-Max standardized. The relevant calculation formula is:

[0112]

[0113] Get the normalized data set:

[0114]

[0115] Among them, C Q* 、E Q* 、N Q* ,I Q* They are normalized consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data.

[0116] In S2, the preprocessing specifically includes first macroscopically observing and comparing the similarity of the change trend of the analyzed load time series with the historical load series, and then analyzing them one by one, focusing on the analysis of the data with mutations, generally from the horizontal (equally spaced moments on the same day) or vertical (same moment on adjacent dates), and finally supplementing the data lost during data collection or storage. The acquired data is preprocessed and analyzed. Since the collected historical data may be distorted or lost, direct use will affect the prediction accuracy, so these data need to be screened and corrected in advance to make the processed data have a certain regularity or make the prediction algorithm insensitive to data changes. Since the collection time of electric vehicles, consumer behavior and renewable energy data may be different, these data must be time-aligned and interpolated, and abnormal data must be corrected, and different data features must be normalized or standardized to ensure that the scales of different features are consistent, so as to facilitate model training. For data lost during data collection or storage, the missing data is supplemented based on the horizontal and vertical correlation of the load.

[0117] S3, performing data dimensionality reduction processing on the normalized data set, then constructing a VMD-GA model, using a genetic algorithm to optimize the parameters of the variational mode decomposition model, and using the data set after dimensionality reduction processing to train the VMD-GA model to obtain a data set after mode decomposition;

[0118] In S3, if the data set is too large, each data set C Q* , EQ* , N Q* , I Q* Perform principal component analysis to reduce dimension. For the data set X∈{C Q* , E Q* , N Q* , I Q* = R N×P , where N is the number of samples and P is the feature dimension. First, center the data X centered =X-μx, μx is the mean vector of each feature, and the covariance matrix is ​​calculated:

[0119]

[0120] Perform eigendecomposition on the covariance matrix:

[0121] ∑{C Q* , E Q* , N Q* , I Q*} i =∑Xν i =λ i ν i

[0122] Among them, v i represents the eigenvector, λ i represents the corresponding eigenvalue, X centered It is a centralized dataset;

[0123] Sort the eigenvectors from large to small according to the eigenvalues ​​of the covariance matrix, select the first h eigenvectors, and form the dimension reduction matrix W h ∈R P×h , and project the data onto the principal components; by Q* , E Q* , N Q* , I Q*} to reduce the dimension, we can get the reduced dimension data set X′:

[0124] X′=X centered W h ∈R N×h ∈{C Q* , E Q* , N Q* , I Q*}′

[0125] The dimension reduction data set X′ is subjected to VMD decomposition, that is, f(t,X′) is taken as input, which is the sum of the data of the distribution network in area e during period t, and is decomposed into m modal components u with specific sparsity m(t), and at the same time determine the center frequency and bandwidth of each modal component. By transforming the data from the time domain to the frequency domain through the variational solution process, the nonlinear time series characteristics of the data can be effectively extracted; it has strong robustness to abnormal points and change points. The variational solution process decomposes the original data into M modal components, minimizing the sum of the estimated bandwidths of each modal component. Its constrained variational model is as follows:

[0126]

[0127] Where M is the number of modal components, z m (t) is the time domain function of the mth subsequence electric power component decomposed from the original electric power, w m is the inherent center frequency of the mth subsequence electric power component, δ(t) is the change rate of the electric power in the decomposed dimensionality reduction data set, j represents the normal complex value of the original electric power, represents the partial derivative with respect to time t;

[0128] In order to obtain the optimal solution of the above variational model, the following formula is used:

[0129]

[0130] Among them, u m (t) represents the time domain function of the mth mode obtained by decomposition, w m represents the center frequency of the mth mode, α is a balance parameter used to control the weight of the bandwidth constraint, Indicates z m (t) is the square of the norm, λ(t) represents the Lagrange multiplication, and <> represents the inner product operation;

[0131] To simplify the calculation, the Lagrangian function is converted to the frequency domain for processing, that is, u m The optimization of (t) is transferred from the time domain to the frequency domain to obtain The frequency domain function of the mth mode obtained by level decomposition; the corresponding z m The frequency domain corresponding to (t) is That is, the frequency domain function of the mth subsequence electric power component decomposed from the original electric power; the frequency domain multiplier corresponding to the λ(t) Lagrange multiplier is Then in the frequency domain:

[0132]

[0133] Then, the alternating multiplier direction algorithm is used to iterate continuously, and the optimal solution can be obtained as follows:

[0134]

[0135] Among them, n is the number of iterations, and τ is the step size parameter for updating the Lagrange multiplier.

[0136] In S3, when extracting features from the preprocessed data, specifically from the consumer behavior dataset C Q It is necessary to extract time features (daily, weekly, monthly, seasonal features, holiday features, distinction between working days and non-working days), temperature and weather features (the relationship between electricity consumption and weather conditions such as temperature and humidity, such as high temperature will lead to increased use of air conditioners), and household equipment usage pattern features (the usage patterns of smart home devices and household appliances, whether there is peak power load); from electric vehicle data E Q The charging characteristics need to be extracted from the data, including charging time period (difference in charging load between daytime and nighttime, concentrated charging period of electric vehicles, location of charging stations), charging amount and charging frequency (average charging amount, charging frequency, and charging duration of each electric vehicle), and the total number of electric vehicles (changing trend of the number of electric vehicles, especially the impact on grid load in specific areas or time periods); from the renewable energy (wind energy, solar energy) dataset N Q The features extracted from the dataset include power generation volatility (solar and wind power generation data, the impact of weather changes on power generation, seasonal changes, especially the difference between winter and summer), renewable energy share (the share of renewable energy in total power supply, especially the fluctuations during peak hours), energy storage system (the use of energy storage system, how to adjust load fluctuations and improve the stability of the power grid); from the digital intelligent data set I Q Extract characteristic real-time load (real-time load data provided by smart meters and sensors, reflecting the electricity demand in each period), demand response (analyzing the load peak and valley differences based on the demand response data provided by the smart grid, and how to adjust the load through dynamic electricity prices), and distributed energy (output data of distributed power generation systems in smart grids, such as the impact of household photovoltaic systems on loads).

[0137] When performing variational modal decomposition on the preprocessed data in S3, local oscillations and periodic behaviors in the signal are extracted, and the complex signal is decomposed into multiple fixed-frequency modal functions, where low-frequency modes are long-term trends, such as consumers' overall electricity consumption behavior patterns, seasonal fluctuations, such as the difference in winter and summer loads, and the impact of holidays on electricity consumption; medium-frequency modes are used to capture fluctuations in charging demand for electric vehicles, such as the peak of electric vehicle charging occurring between 7 and 9 p.m., especially household charging behavior; high-frequency modes are used to capture short-term load fluctuations, such as load fluctuations of zero caused by weather changes in a short period of time, or local fluctuations at electric vehicle charging stations due to high loads;

[0138] The genetic algorithm is used to optimize the parameters of the variational mode decomposition model. In the genetic algorithm, for C Q′ 、EQ′ 、N Q′ ,I Q′ The fitness of each data set can be expressed as follows:

[0139]

[0140] Among them, u Cm (t),u Em (t),u Nm (t),u Im (t) are C Q′ 、E Q′ 、N Q′ ,I Q′ The modes of the data set obtained by VMD decomposition; the optimal mode number after genetic algorithm optimization is as follows:

[0141]

[0142] Among them, M C ,M E ,M N ,M I represents the optimal number of modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) represents the best modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) together constitute the data set after modal decomposition.

[0143] The genetic algorithm (GA) in S3 is based on the theory of biological evolution. It changes the search process through selection, crossover and mutation to find the optimal solution. It is used to optimize the weights and thresholds of the BP neural network to solve the problem of being trapped in the local optimal solution, while speeding up the training speed of the model, enhancing the adaptability and convergence of the model, thereby improving the recognition accuracy.

[0144] S4. Construct the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model. The IPSO-GRU model is trained using the data set after modal decomposition.

[0145] The particle swarm algorithm is used to optimize the hyperparameters of the GRU model:

[0146] Set some hyperparameters of GRU, initialize IPSO parameters, find the optimal solution through IPSO iteration, input the optimized hyperparameters into GRU model training, and output the predicted value. Its mathematical description is shown in the following formula;

[0147] z t =σ(W z ·[h t-1 ,x t ])

[0148] r t =σ(W r ·[h t-1 ,x t ])

[0149] h t =φ(W·[r t *h t-1 ,x t ])

[0150] h t =(1-z t )*h t-1 +z t *h t

[0151] Among them, x t is the input vector at time step t, using the data set C after modal decomposition Q (t), E Q (t), N Q (t), I Q (t) as the input vector, h t-1 and h t are the state variables of the previous moment and this moment respectively, r t is the update gate, z t is the reset gate, h t is a candidate set; W r , W z , W h is the weight parameter; σ is the Sigmoid activation function; φ is the Tanh activation function;

[0152] The mathematical descriptions of σ and φ are as follows:

[0153]

[0154] The particle swarm algorithm is used to optimize some hyperparameters of the GRU model. Q′ 、E Q′ 、N Q′ ,I Q′ Data set, output prediction value as follows:

[0155]

[0156]

[0157] in, All are bias items. After fusing the data sets, the predicted values ​​can be output:

[0158]

[0159] An IPSO-GRU model based on the LSTM model is established. In terms of consumer behavior, GRU can capture changes in consumer electricity consumption patterns through its time series modeling capabilities, such as the difference in electricity demand between daytime and nighttime, or the difference in electricity consumption between holidays and weekdays; in terms of electric vehicle charging, by learning from historical electric vehicle charging data, GRU can capture the peak charging period of electric vehicles, and then accurately predict the impact of charging on the grid load; in terms of renewable energy fluctuations, GRU can effectively handle the fluctuations in renewable energy generation such as solar and wind power, and combine weather data (such as wind speed and sunshine duration) to help the model capture the periodicity and randomness in load changes; therefore, through the GRU gating mechanism, the long-term trends (such as seasonal changes) and short-term fluctuations (such as the charging peak of electric vehicles) in the grid load data can be effectively captured, and the gradient vanishing problem can be avoided while considering time dependence, thereby improving the prediction accuracy.

[0160] In the process of using IPSO to optimize the GRU neural network in S4, some hyperparameters of the GRU neural network (such as network nodes, number of hidden layers, and batch size, etc.) are first fixed, and the number of neurons m and time step T are input into the IPSO algorithm as parameters to be optimized. Then, the population size of the particle swarm, the maximum and minimum values ​​of the particle velocity, the learning rate, the maximum number of iterations, and the initial position of the particles are initialized. Finally, the IPSO algorithm is used for iterative optimization to find and record the optimal solution and update the optimal position of the particles. When the trained parameters meet the set conditions, the iteration is stopped and the optimal parameter value is output; otherwise, the training is continued. Finally, the hyperparameters optimized by the IPSO algorithm are used as the initial hyperparameters of the GRU neural network for training, and the predicted values ​​are output.

[0161] The IPSO-GRU model in S4 is a variant of the LSTM model. The GRU neural network retains the memory function of the LSTM network and can effectively process the long-term relationship between sequence data. The update gate of the GRU neural network can greatly improve the speed of data processing and has autonomous learning capabilities.

[0162] The GRU neural network model in S4 consists of three parts: input layer, hidden layer and output layer. The forget gate and input gates are combined in an update gate. The update gate contains both neuron state and hidden state, which can reduce the complexity of network units, reduce the number of parameters, and greatly shorten the training time of the model. The function of the update gate is to limit how much data state of the previous neural unit is saved to the current unit. The larger the value of the update gate, the more data state of the previous neural unit is saved. The reset gate is used to limit how much data state of the previous neural unit is input to the current neural unit. The larger the value of the reset gate, the more data of the previous unit is written.

[0163] S5. Collect the operating parameters of the target area and make power load forecasts for the next target period based on the VMD-GA model and the IPSO-GRU model.

[0164] The VMD-GA-IPSO-GRU based short-term power load forecasting method proposed in S5 analyzes the final forecast results, outputs the forecast values, compares the fitting curves of the forecast data with the test data, and evaluates the forecast model.

[0165] The target area for load forecasting is a certain area in the south in 2018. Consumer behavior data, electric vehicle data, energy transformation related data, and digital intelligence data are collected within two days. The data interval is 15 minutes to form a raw data set. The collected data is preprocessed and decomposed according to VMD-GA, such as Figure 1 This is the decomposition prediction diagram of the preprocessed data.

[0166] Furthermore, the preprocessed data is input into the VMD-GA-IPSO-GRU model, and a data is measured every 15 minutes, with a total of 200 measurement points. It can be concluded that the VMD-GA-IPSO-GRU model can well predict the short-term power load under the new economic situation (predicted value), such as Figure 2 As shown. Among them, the population size of PSO is set to 50, the number of mixed iterations is 500, the inertia weight is 0.8, the network parameters of the initial neural network have 1 input layer, 1 output layer, 3 hidden layers, the range of the number of neurons is [50,300], the range of the time step is [1,5], the minimum number of training is 1, the learning rate is 0.001, the learning rate reduction method is adam, the maximum number of iterations is 250, and the number of iterations per round is 1. The prediction results of the present invention are compared with the real data, and the traditional GRU prediction model is selected as the comparison value. The simulation environment is MATLAB, as shown in Figure 2Obviously, compared with the traditional GRU prediction method, the VMD-GA-IPSO-GRU model proposed in this patent has higher prediction accuracy, and the relative error is reduced from 3.9% to 1.2%, which can improve the prediction accuracy by 2.7% compared with GRU.

[0167] Embodiment 2:

[0168] Embodiment 2 is substantially the same as Embodiment 1, except that:

[0169] In the training of VMD-GA model and IPSO-GRU model, if there is still large-scale data calculation after data dimension reduction, the large-scale data set can be divided into smaller batches through small batch training. Only one batch of data is used for calculation each time training, which reduces memory usage and speeds up training. The data set to be trained is divided into N batches. batch Batches:

[0170]

[0171] The amount of training for each batch is:

[0172]

[0173] Where B is the batch size.

[0174] Embodiment 3:

[0175] See also Figure 2 , a short-term power load forecasting system under the new economic situation, the system is used to execute the short-term power load forecasting method under the new economic situation, specifically including: a data acquisition module, a data preprocessing module, a VMD-GA model building module, an IPSO-GRU model building module, and a power load forecasting module;

[0176] The data collection module is used to determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a collection of original data;

[0177] Determine the load forecast target area e and form an original data set;

[0178] The original data set includes:

[0179] C Q =C Q 1 +C Q 2 +C Q 3 +C Q 4 +CQ 5

[0180] E Q =E Q 1 +E Q 2 +E Q 3 +E Q 4 +E Q 5

[0181] N Q =N Q 1 +N Q 2 +N Q 3 +N Q 4

[0182] I Q =I Q 1 +I Q 2 +I Q 3 +I Q 4

[0183] Where: C Q For consumer behavior data, C Q 1 is the industrial electricity consumption data, C Q 2 is household electricity consumption data, C Q 3 is the commercial electricity consumption data, C Q 4 The usage pattern of home devices on the cloud platform, C Q 5 is the user behavior history data, E Q For electric vehicle data, E Q 1 For electric vehicle charging pile data, E Q 2 Charging behavior data for electric vehicles, E Q 3 is the charging period data, E Q 4 Charging capacity for electric vehicles, E Q 5 is the location data of the charging station, N Q For data related to energy transformation, N Q 1is the wind power generation data, N Q 2 is the solar power generation data, N Q 3 is the energy policy related data, N Q 4 is the energy storage system data, I Q To digitize intelligent data, I Q 1 For smart meter data, I Q 2 For sensor and IoT data, I Q 3 Real-time power consumption data for IoT devices, Q 4 For cloud computing platform data.

[0184] The data preprocessing module is used to perform preprocessing and analysis on the data in the original data set, and then perform normalization processing to obtain a normalized data set;

[0185] Preprocess and analyze the data in the original data set, and fill in the missing or abnormal data according to the following formula:

[0186] C Q (d,t)=a c1 C Q (d-1,t)+a c2 C Q (d+1,t)+a c3 C Q (d,t-1)+a c4 C Q (d,t+1)

[0187] E Q (d,t)=a e1 E Q (d-1,t)+a e2 E Q (d+1,t)+a e3 E Q (d,t-1)+a e4 E Q (d,t+1)

[0188] N Q (d,t)=a n1 N Q (d-1,t)+a n2 N Q (d+1,t)+a n3 N Q (d,t-1)+a n4 N Q(d,t+1)

[0189] I Q (d,t)=a i1 I Q (d-1,t)+a i2 I Q (d+1,t)+a i3 I Q (d,t-1)+a i4 I Q (d,t+1)

[0190] Among them, C Q (d,t),E Q (d,t),N Q (d,t),I Q (d, t) represent the consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data at the t-th time on the d-th day that need to be corrected; a c1 、a c2 、a c3 、a c4 are the weight coefficients for completing consumer behavior data; a e1 、a e2 、a e3 、a e4 are weight coefficients for completing data related to energy transformation; a n1 、a n2 、a n3 、a n4 are the weight coefficients for electric vehicle data completion; a i1 、a i2 、a i3 、a i4 They are all weight coefficients for digital and intelligent data completion;

[0191] Normalize the preprocessed data to get the normalized data set:

[0192]

[0193] Among them, C Q* 、E Q* 、N Q* ,I Q* They are normalized consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data.

[0194] The VMD-GA model building module is used to perform data dimensionality reduction processing on the normalized data set, then build the VMD-GA model, use the genetic algorithm to optimize the parameters of the variational mode decomposition model, and use the data set after dimensionality reduction processing to train the VMD-GA model to obtain the data set after mode decomposition;

[0195] First, for each data set C Q* 、E Q* 、N Q* ,I Q* Perform principal component analysis to reduce dimension. For the data set X∈{C Q* , E Q* , N Q* , I Q* = R N×P , where N is the number of samples and P is the feature dimension. First, center the data X centered =X-μx, μx is the mean vector of each feature, and the covariance matrix is ​​calculated:

[0196]

[0197] Perform eigendecomposition on the covariance matrix:

[0198] ∑{C Q* , E Q* , N Q* , I Q*} i =∑Xν i =λ i ν i

[0199] Among them, v i represents the eigenvector, λ i represents the corresponding eigenvalue, X centered It is a centralized dataset;

[0200] Sort the eigenvectors from large to small according to the eigenvalues ​​of the covariance matrix, select the first h eigenvectors, and form the dimension reduction matrix W h ∈R P×h , and project the data onto the principal components; by Q* , E Q* , N Q* , I Q*} to reduce the dimension, we can get the reduced dimension data set X′:

[0201] X′=X centered W h ∈R N×h ∈{C Q* , E Q* , N Q* , I Q*}′

[0202] The dimension reduction data set X′ is subjected to VMD decomposition, that is, f(t,X′) is taken as input, which is the sum of the data of the distribution network in area e during period t, and is decomposed into m modal components u with specific sparsity m (t), and determine the center frequency and bandwidth of each modal component at the same time. The nonlinear time series characteristics of the data can be effectively extracted by the variational solution process that transforms the data from the time domain to the frequency domain. The variational solution process decomposes the original data into M modal components, minimizing the sum of the estimated bandwidths of each modal component. The constrained variational model is as follows:

[0203]

[0204] Where M is the number of modal components, z m (t) is the time domain function of the mth subsequence electric power component decomposed from the original electric power, w m is the inherent center frequency of the mth subsequence electric power component, δ(t) is the change rate of the electric power in the decomposed dimensionality reduction data set, j represents the normal complex value of the original electric power, represents the partial derivative with respect to time t;

[0205] In order to obtain the optimal solution of the above variational model, the following formula is used:

[0206]

[0207] Among them, u m (t) represents the time domain function of the mth mode obtained by decomposition, w m represents the center frequency of the mth mode, α is a balance parameter used to control the weight of the bandwidth constraint, Indicates z m (t) is the square of the norm, λ(t) represents the Lagrange multiplication, and <> represents the inner product operation;

[0208] To simplify the calculation, the Lagrangian function is converted to the frequency domain for processing, that is, u m The optimization of (t) is transferred from the time domain to the frequency domain to obtain The frequency domain function of the mth mode obtained by level decomposition; the corresponding z m The frequency domain corresponding to (t) is That is, the frequency domain function of the mth subsequence electric power component decomposed from the original electric power; the frequency domain multiplier corresponding to the λ(t) Lagrange multiplier is Then in the frequency domain:

[0209]

[0210] Then, the alternating multiplier direction algorithm is used to iterate continuously, and the optimal solution can be obtained as follows:

[0211]

[0212] Among them, n is the number of iterations, and τ is the step size parameter for updating the Lagrange multiplier.

[0213] The genetic algorithm is used to optimize the parameters of the variational mode decomposition model. In the genetic algorithm, for C Q′ 、E Q′ 、N Q′ ,I Q′ The fitness of each data set can be expressed as follows:

[0214]

[0215] Among them, u Cm (t),u Em (t),u Nm (t),u Im (t) are C Q′ 、E Q′ 、N Q′ ,I Q′ The modes of the data set obtained by VMD decomposition; the optimal mode number after genetic algorithm optimization is as follows:

[0216]

[0217] Among them, M C ,M E ,M N ,M I represents the optimal number of modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) represents the best modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) together constitute the data set after modal decomposition.

[0218] The IPSO-GRU model building module is used to build the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model and train the IPSO-GRU model using the data set after modal decomposition.

[0219] The particle swarm algorithm is used to optimize the hyperparameters of the GRU model:

[0220] Set some hyperparameters of GRU, initialize IPSO parameters, find the optimal solution through IPSO iteration, input the optimized hyperparameters into GRU model training, and output the predicted value. Its mathematical description is shown in the following formula;

[0221] z t =σ(W z ·[h t-1 ,x t ])

[0222] r t =σ(W r ·[h t-1 ,x t ])

[0223] h t =φ(W·[r t *h t-1 ,x t ])

[0224] h t =(1-z t )*h t-1 +z t *h t

[0225] Among them, x t is the input vector at time step t, using the data set C after modal decomposition Q (t), E Q (t), N Q (t), I Q (t) as the input vector, h t-1 and h t are the state variables of the previous moment and this moment respectively, r t is the update gate, z t is the reset gate, h t is a candidate set; W r , W z , W h is the weight parameter; σ is the Sigmoid activation function; φ is the Tanh activation function;

[0226] Among them, the mathematical descriptions of σ and φ are as follows:

[0227]

[0228] The particle swarm algorithm is used to optimize some hyperparameters of the GRU model. Q′ 、E Q′ 、N Q′ ,I Q′Data set, output prediction value as follows:

[0229]

[0230] in, All are bias items, which fuse the data sets:

[0231]

[0232] Output of short-term power load forecasts

[0233] The power load forecasting module is used to collect the operating parameters of the target area and forecast the power load for the next target period based on the VMD-GA model and the IPSO-GRU model.

[0234] Embodiment 4:

[0235] See also Figure 3 A short-term power load forecasting device under the new economic situation includes a memory and a processor, wherein the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the above-mentioned short-term power load forecasting method under the new economic situation according to the instructions in the computer program code.

[0236] Embodiment 5:

[0237] A computer program product comprises a computer program, wherein the computer program is executed by a processor to implement the short-term power load forecasting method under the new economic situation.

[0238] Embodiment 6:

[0239] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the short-term power load forecasting method under the above-mentioned new economic situation.

Claims

1. A short-term power load forecasting method under the new economic situation, characterized in that: The steps include: S1. Determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a raw data set; S2, after preprocessing and analyzing the data in the original data set, normalization is performed to obtain a normalized data set; S3, performing data dimensionality reduction processing on the normalized data set, then constructing a VMD-GA model, using a genetic algorithm to optimize the parameters of the variational mode decomposition model, and using the data set after dimensionality reduction processing to train the VMD-GA model to obtain a data set after mode decomposition; S4. Construct the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model. The IPSO-GRU model is trained using the data set after modal decomposition. S5. Collect the operating parameters of the target area and make power load forecasts for the next target period based on the VMD-GA model and the IPSO-GRU model.

2. The method for short-term power load forecasting under the new economic situation according to claim 1 is characterized by: In said S1, a load forecast target area e is determined to form an original data set; The original data set includes: C Q =C Q 1+C Q 2+C Q 3+C Q 4+C Q 5 AND Q =And Q 1+E Q 2+E Q 3+E Q 4+E Q 5 N Q =N Q 1+N Q 2+N Q 3+N Q 4 I Q =I Q 1+I Q 2+I Q 3+I Q 4 Where: C Q For consumer behavior data, C Q 1 is industrial electricity consumption data, C Q 2 is household electricity consumption data, C Q 3 is the commercial electricity consumption data, C Q 4 is the usage mode of home devices on the cloud platform, C Q 5 is the user behavior history data, E Q For electric vehicle data, E Q 1 is the data of electric vehicle charging pile, E Q 2 is the charging behavior data of electric vehicles, E Q 3 is the charging period data, E Q 4 is the charging capacity of the electric vehicle, E Q 5 is the location data of the charging station, N Q For data related to energy transformation, N Q 1 is wind power generation data, N Q 2 is the solar power generation data, N Q 3 is energy policy related data, N Q 4 is the energy storage system data, I Q To digitize intelligent data, I Q 1 is the smart meter data, I Q 2 is sensor and IoT data, I Q 3 is the real-time power consumption data of IoT devices, Q 4 is the cloud computing platform data.

3. The short-term power load forecasting method under the new economic situation according to claim 1 is characterized by: In S2, the data in the original data set is preprocessed and analyzed, and missing data or abnormal data are supplemented according to the following formula: C Q (d,t)=a c1 C Q (d-1,t)+a c2 C Q (d+1,t)+a c3 C Q (d,t-1)+a c4 C Q (d,t+1) E Q (d,t)=a e1 E Q (d-1,t)+a e2 E Q (d+1,t)+a e3 E Q (d,t-1)+a e4 E Q (d,t+1) N Q (d,t)=a n1 N Q (d-1,t)+a n2 N Q (d+1,t)+a n3 N Q (d,t-1)+a n4 N Q (d,t+1) I Q (d,t)=a i1 I Q (d-1,t)+a i2 I Q (d+1,t)+a i3 I Q (d,t-1)+a i4 I Q (d,t+1) Among them, C Q (d,t),E Q (d,t),N Q (d,t),I Q (d, t) represent the consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data at the t-th time on the d-th day that need to be corrected; a c1 、a c2 、a c3 、a c4 are the weight coefficients for completing consumer behavior data; a e1 、a e2 、a e3 、a e4 are weight coefficients for completing data related to energy transformation; a n1 、a n2 、a n3 、a n4 are the weight coefficients for electric vehicle data completion; a i1 、a i2 、a i3 、a i4 They are all weight coefficients for digital intelligent data completion; Normalize the preprocessed data to get the normalized data set: Among them, C Q* 、E Q* 、N Q* ,I Q* They are normalized consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data.

4. The method for short-term power load forecasting under the new economic situation according to claim 1 is characterized in that: In S3, first, each data set C Q* , E Q* , N Q* , I Q* Perform principal component analysis to reduce dimension. For the data set X∈{C Q* , E Q* , N Q* , I Q* = R N×P , where N is the number of samples and P is the feature dimension. First, center the data X centered =X-μx, μx is the mean vector of each feature, and the covariance matrix is ​​calculated: Perform eigendecomposition on the covariance matrix: ∑ {C Q* ,E Q* ,N Q* ,I Q* }n i =∑Xν i =λ i n i Among them, v i represents the eigenvector, λ i represents the corresponding eigenvalue, X centered It is a centralized dataset; Sort the eigenvectors from large to small according to the eigenvalues ​​of the covariance matrix, select the first h eigenvectors, and form the dimension reduction matrix W h ∈R P×h , and project the data onto the principal components; by Q* , E Q* , N Q* , I Q* } to reduce the dimension, we can get the reduced dimension data set X′: X′=X centered W h ∈R N×h ∈{C Q* ,E Q* ,N Q* ,I Q* }′ The dimension reduction data set X′ is subjected to VMD decomposition, that is, f(t,X′) is taken as input, which is the sum of the data of the distribution network in area e during period t, and is decomposed into m modal components u with specific sparsity m (t), and determine the center frequency and bandwidth of each modal component at the same time. The nonlinear time series characteristics of the data can be effectively extracted by the variational solution process that transforms the data from the time domain to the frequency domain. The variational solution process decomposes the original data into M modal components, minimizing the sum of the estimated bandwidths of each modal component. The constrained variational model is as follows: Where M is the number of modal components, z m (t) is the time domain function of the mth subsequence electric power component decomposed from the original electric power, w m is the inherent center frequency of the mth subsequence electric power component, δ(t) is the change rate of the electric power in the decomposed dimensionality reduction data set, j represents the normal complex value of the original electric power, represents the partial derivative with respect to time t; In order to obtain the optimal solution of the above variational model, the following formula is used: Among them, u m (t) represents the time domain function of the mth mode obtained by decomposition, w m represents the center frequency of the mth mode, α is a balance parameter used to control the weight of the bandwidth constraint, Indicates z m (t) is the square of the norm, λ(t) represents the Lagrange multiplication, and <> represents the inner product operation; To simplify the calculation, the Lagrangian function is converted to the frequency domain for processing, that is, u m The optimization of (t) is transferred from the time domain to the frequency domain to obtain The frequency domain function of the mth mode obtained by level decomposition; the corresponding z m The frequency domain corresponding to (t) is That is, the frequency domain function of the mth subsequence electric power component decomposed from the original electric power; the frequency domain multiplier corresponding to the λ(t) Lagrange multiplier is Then in the frequency domain: Then, the alternating multiplier direction algorithm is used to iterate continuously, and the optimal solution can be obtained as follows: Among them, n is the number of iterations, and τ is the step size parameter for updating the Lagrange multiplier.

5. The method for short-term power load forecasting under the new economic situation according to claim 3 is characterized by: In S3, a genetic algorithm is used to optimize the parameters of the variational mode decomposition model. In the genetic algorithm, for C Q′ 、E Q ′、N Q′ ,I Q The fitness of each data set can be expressed as follows: Among them, u Cm (t),u Em (t),u Nm (t),u Im (t) are C Q′ 、E Q′ 、N Q′ ,I Q′ The modes of the data set obtained by VMD decomposition; the optimal mode number after genetic algorithm optimization is as follows: Among them, M C ,M E ,M N ,M I represents the optimal number of modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) represents the best modes obtained by GA-VMD optimization of consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data; C Q (t), E Q (t), N Q (t), I Q (t) together constitute the data set after modal decomposition.

6. The method for short-term power load forecasting under the new economic situation according to claim 1 is characterized by: In S4, the particle swarm algorithm is used to optimize the hyperparameters of the GRU model: Set some hyperparameters of GRU, initialize IPSO parameters, find the optimal solution through IPSO iteration, input the optimized hyperparameters into GRU model training, and output the predicted value. Its mathematical description is shown in the following formula; z t =σ(W z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) h t =φ(W·[r t *h t-1 ,x t ]) h t =(1-z t )*h t-1 +z t *h t Among them, x t is the input vector at time step t, using the data set C after modal decomposition Q (t), E Q (t), N Q (t), I Q (t) as the input vector, h t-1 and h t are the state variables of the previous moment and this moment respectively, r t is the update gate, z t is the reset gate, h t is a candidate set; W r , W z , W h is the weight parameter; σ is the Sigmoid activation function; φ is the Tanh activation function; Among them, the mathematical descriptions of σ and φ are as follows: The particle swarm algorithm is used to optimize some hyperparameters of the GRU model. Q′ 、E Q′ 、N Q′ ,I Q′ Data set, output prediction value as follows: in, All are bias items, which fuse the data sets: Output of short-term power load forecasts 7. A short-term power load forecasting system under the new economic situation, characterized in that: The system is used to execute the short-term power load forecasting method under the new economic situation as described in any one of claims 1 to 6, specifically comprising: a data acquisition module, a data preprocessing module, a VMD-GA model building module, an IPSO-GRU model building module, and a power load forecasting module; The data collection module is used to determine the target area for load forecasting, collect consumer behavior data, electric vehicle data, energy transformation-related data, and digital intelligence data to form a collection of original data; The data preprocessing module is used to perform preprocessing and analysis on the data in the original data set, and then perform normalization processing to obtain a normalized data set; The VMD-GA model building module is used to perform data dimensionality reduction processing on the normalized data set, then build the VMD-GA model, use the genetic algorithm to optimize the parameters of the variational mode decomposition model, and use the data set after dimensionality reduction processing to train the VMD-GA model to obtain the data set after mode decomposition; The IPSO-GRU model building module is used to build the IPSO-GRU model. The particle swarm algorithm is used to optimize the hyperparameters of the GRU model and train the IPSO-GRU model using the data set after modal decomposition. The power load forecasting module is used to collect the operating parameters of the target area and forecast the power load for the next target period based on the VMD-GA model and the IPSO-GRU model.

8. A short-term power load forecasting device under the new economic situation, characterized in that: The method comprises a memory and a processor, wherein the memory is used to store computer program code and transmit the computer program code to the processor; The processor is used to execute the short-term power load forecasting method under the new economic situation as described in any one of claims 1 to 5 according to the instructions in the computer program code.

9. A computer program product, comprising a computer program, characterized in that The computer program is executed by a processor to implement the short-term power load forecasting method under the new economic situation as described in any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the short-term power load forecasting method under the new economic situation as described in any one of claims 1 to 6.

Images