Virtual power plant load prediction method and system based on multi-model combination

Abstract

The invention discloses a virtual power plant load prediction method and system based on multi-model combination, and relates to the technical field of virtual power plants, and the method comprises the steps: carrying out the preprocessing of obtained multi-source heterogeneous data; clustering the preprocessed historical load data, and dividing the historical load data into at least two load feature groups according to different seasons or power utilization modes; respectively utilizing the trained linear regression model, the trained LSTM neural network model and the trained support vector machine model to correspondingly obtain three preliminary prediction results; determining a weight coefficient of the model by using a genetic algorithm to obtain a combined prediction result; and carrying out reverse normalization processing on the combined prediction result, and outputting a load prediction value of the virtual power plant in a future preset time period. The problem that a traditional single model is poor in adaptability to complex relevance, dynamic change and external interference is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of virtual power plants, in particular to a virtual power plant load prediction method and system based on multi-model combination. BACKGROUND

[0002] The virtual power plant load refers to the prediction and management of power demand of the virtual power plant in the operation process, and the core is to integrate distributed energy resources and realize dynamic scheduling to ensure the stability and economy of the power system. The virtual power plant load prediction involves comprehensive analysis of multi-dimensional information such as historical data, weather factors, price fluctuations and social events, and is a key link for optimizing energy allocation and improving response efficiency.

[0003] In the prior art, the single model used for virtual power plant load prediction mainly relies on linear regression model, LSTM neural network model, support vector machine model or traditional statistical model, etc. independent algorithm, and the core problem is that the model structure is fixed and the parameter optimization is single, which is difficult to cope with the complex correlation and dynamic change characteristics of multi-source heterogeneous data. Among them, the linear regression model and the nonlinear fluctuation have insufficient adaptability. The linear regression model can only fit the linear correlation between the characteristics and the load, and cannot capture the nonlinear fluctuations caused by extreme weather (such as air conditioning load surge caused by high temperature) and price mutation, resulting in significant increase of prediction error. The LSTM neural network model is sensitive to data noise, and when there are abnormal values (such as load jump caused by sensor failure) in the input data, overfitting phenomenon is easy to occur, which cannot be generalized to new power consumption scenarios; the support vector machine model is prone to "dimension disaster" when the feature dimension is too high (such as more than 10 features including weather, price, green certificate transaction, etc.), which leads to decline of model training efficiency and prediction accuracy.

[0004] And the statistical model such as ARIMA cannot adapt to sudden events or long-term trend changes; it lacks multi-model coordination mechanism and dynamic adjustment ability, resulting in significant defects in prediction result accuracy stability, external disturbance adaptability and long-term operation reliability, which cannot meet the demand of virtual power plant for high real-time and high robustness of load prediction SUMMARY In view of the deficiencies in the prior art, the present application provides a virtual power plant load prediction method and system based on multi-model combination, which solves the problem of poor adaptability of traditional single model to complex correlation, dynamic change and external disturbance.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows: In a first aspect, the present application provides a virtual power plant load prediction method based on multi-model combination, comprising: obtaining multi-source heterogeneous data containing virtual power plant historical load data, and preprocessing the multi-source heterogeneous data; clustering the pretreated historical load data, and dividing at least two load feature groups according to different seasons or power consumption modes according to a clustering result; based on the load feature groups, respectively using a trained linear regression model, an LSTM neural network model and a support vector machine model to obtain three preliminary prediction results correspondingly; based on historical prediction errors of the three preliminary prediction results correspondingly, using a genetic algorithm to determine weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model, and obtaining a combined prediction result; performing inverse normalization processing on the combined prediction result, and outputting a load prediction value of the virtual power plant in a future preset time period.

[0006] As a further technical solution, the multi-source heterogeneous data further includes weather data, electricity price data, green certificate transaction data and major event data of the virtual power plant; the preprocessing includes timestamp alignment, outlier correction, normalization processing and feature screening, wherein the timestamp alignment takes the historical load data timestamp as a reference; the outlier correction adopts horizontal comparison method and vertical comparison method; the normalization processing adopts min-max normalization; and the feature screening adopts Pearson correlation coefficient to calculate the correlation of the weather data, the electricity price data, the green certificate transaction data, the major event data and the historical load data respectively.

[0007] As a further technical solution, the linear regression model adopts a multivariate linear regression model, and is expressed as: ; wherein, is an input feature vector, is a weight parameter, is a bias term, is the first preliminary prediction result of the output.

[0008] As a further technical solution, the cell state update formula of the LSTM unit in the LSTM neural network model is: ; The formula of the forget gate output, the input gate output and the current time hidden state is: ; ; ; wherein, is the current time cell state, is the previous time cell state, is the forget gate output, is the input gate output, is the candidate cell state, Hide the current state. This is the hidden state from the previous moment. Enter the current time. This is the weight matrix. This is a bias term.

[0009] As a further technical solution, the support vector machine model adopts... The objective function of the support vector regression model is: ;in, Weight vector, As slack variables, The penalty coefficient is... For sample size; Its constraints are: ;in, For the first One input feature vector, For kernel function mapping, For bias terms, For the target value, This is an insensitive loss parameter.

[0010] As a further technical solution, the historical prediction error of the model corresponding to the three preliminary prediction results is calculated using the symmetrical mean absolute percentage error, and expressed as follows: ;in, For predicted values, This is the actual value. To predict the sample size.

[0011] As a further technical solution, a genetic algorithm is used to determine the weight coefficients of the linear regression model, the LSTM neural network model, and the support vector machine model. Specifically, real-number encoding is used, with each chromosome corresponding to a set of weight coefficients. ,satisfy + + =1 and ≥0; The fitness function of the genetic algorithm uses the symmetric average absolute percentage error value of the combined prediction results as the fitness value, and the selection operator of the genetic algorithm adopts tournament selection, the crossover operator adopts arithmetic crossover, and the mutation operator adopts uniform mutation.

[0012] Secondly, the present invention provides a virtual power plant load forecasting system based on multi-model combination, comprising the following modules: The data acquisition and preprocessing module is configured to: acquire multi-source heterogeneous data containing historical load data of virtual power plants, and preprocess the multi-source heterogeneous data; The clustering module is configured to cluster the preprocessed historical load data, and divide at least two load feature groups according to different seasons or power consumption modes according to the clustering result; The preliminary prediction module is configured to obtain three preliminary prediction results by respectively using the trained linear regression model, the LSTM neural network model and the support vector machine model based on the load feature groups; The weighting module is configured to determine the weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model by using a genetic algorithm based on the historical prediction errors of the three preliminary prediction results corresponding models, and obtain a combined prediction result; The output module is configured to perform inverse normalization processing on the combined prediction result, and output the load prediction value of the virtual power plant in the future preset time period.

[0013] The one or more technical solutions of the present application have the following beneficial effects: The present application solves the limitation of a single model in the prior art by using multiple model cooperation and dynamic weighted combination. The present application uses linear regression model, LSTM neural network model and support vector machine model for preliminary prediction by using each load feature group. Linear regression captures linear relationship, LSTM processes time series nonlinear fluctuation, and support vector machine model enhances high-dimensional feature generalization ability. Then, the genetic algorithm is used to optimize the weight coefficients of each model, and the combined result is obtained by weighted summation, which integrates the advantages of each model, fully captures the load characteristics, and makes up for the shortcomings of single model in processing linear and nonlinear, high-dimensional features, etc.

[0014] The present application solves the problem of adapting to multiple power consumption modes by dividing the load feature groups by the DBSCAN clustering algorithm. The preprocessed historical load data is clustered, and groups corresponding to different seasons or power consumption modes are divided. Each group independently uses multiple model prediction, so that the model can be adjusted according to the characteristics of different power consumption modes, avoiding the defects of single model that is difficult to adapt to the change of multiple loads due to fixed structure, and improving the adaptability to different scenes.

[0015] After obtaining the multi-source heterogeneous data, the present application aligns the time stamp to unify the granularity, corrects the abnormal value, corrects the abnormality by horizontal comparison and vertical interpolation, normalizes the standard data, selects high-correlation features, reduces the noise and redundancy of multi-source heterogeneous data, provides reliable input for subsequent models, and alleviates the overfitting problem caused by noise interference of LSTM and other models. BRIEF DESCRIPTION OF DRAWINGS

[0016] The drawings accompanying the specification of the present application form part of the present application and serve to provide a further understanding of the present application. The illustrative embodiments of the present application and their description serve to explain the present application without unduly limiting it.

[0017] Figure 1 A flow chart of the virtual power plant load prediction method based on multi-model combination in the present application; Figure 2 A system architecture diagram of the virtual power plant load prediction system based on multi-model combination in the present application. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present application have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0019] Embodiment one The present embodiment is applied to a virtual power plant covering distributed wind power, photovoltaic, energy storage, and industrial and residential loads, and needs to realize 72-hour load prediction in the future to support dispatching decisions.

[0020] A virtual power plant load prediction method based on multi-model combination is provided in the present embodiment, as shown in Figure 1 The specific method steps are as follows: S1: Obtain multi-source heterogeneous data including historical load data, weather data, electricity price data, green certificate transaction data and major event data of the virtual power plant, and pre-process the multi-source heterogeneous data.

[0021] In step S1, the historical load data is 2020-2022 hourly load data, with a unit of megawatt, a total of 26280, and the data structure refers to the historical data format of similar energy systems. The weather data is the temperature, humidity, wind speed and light intensity in the same period, with units of degrees Celsius, percentage, meters per second and watts per square meter, respectively, and the collection frequency is synchronized with the historical load data. The electricity price data covers time-of-use electricity price and real-time market electricity price, and the time-of-use electricity price is divided into peak segment 0.8 yuan per kilowatt-hour, flat segment 0.5 yuan per kilowatt-hour and valley segment 0.3 yuan per kilowatt-hour, which is consistent with the conventional price mechanism of the power market. The green certificate transaction data uses the monthly moving average of the non-synchronous generation instantaneous penetration rate, and the Pearson correlation coefficient between this index and the load is 0.41, which belongs to moderate correlation, and is consistent with the typical results in similar analysis. The major event data includes statutory holidays, regional power outage maintenance records, extreme weather warnings, etc., and all data are collected once an hour.

[0022] In step S1, the pre-processing includes timestamp alignment, outlier correction, normalization processing and feature selection, specifically: Timestamp alignment is based on the timestamps of historical load data, such as 00:00 on January 1, 2022. Asynchronously collected meteorological data (once every 10 minutes) is resampled to the hour level using linear interpolation. In practice, six 10-minute data points are taken within a certain hour, and the average meteorological value for that hour is calculated through linear fitting. The calculation method is that the hourly meteorological value is equal to the sum of the products of each 10-minute data point within that hour and its corresponding weight. The weights are allocated according to the time proportion to ensure that the time granularity of all data is unified to the hour level.

[0023] Outlier correction employs a combination of horizontal and vertical comparison methods. The horizontal comparison method calculates the average load data from four time points before and after a given moment as a baseline. The calculation method is as follows: the baseline value at time t on day m is equal to the sum of the load values ​​at times t-2, t-1, t+1, and t+2 on day m, divided by 4. When the load value at that moment exceeds 1.2 times the baseline value or falls below 0.8 times, the baseline value is used instead. This threshold is set with reference to empirical values ​​for outlier identification in similar studies. The formula for calculating the baseline value is: ;in, For the first sky t The baseline value at any given time. For the first sky t+j The load value at a given time; if the load value at that time exceeds 1.2 times or is less than 0.8 times the baseline value, replace it with the average value.

[0024] The longitudinal comparison method compares the load difference at the same time on adjacent dates. When the difference exceeds 1.5 times the historical average difference, linear interpolation is used for correction. The correction formula is: the corrected load value at time t+1 on day m equals the load value at time t+1 on day (m-1) plus the load difference between day m and time t on day (m-1). This method is consistent with the principle of conventional data longitudinal correction and is expressed as follows: ; in, For the first sky t Corrected load value at time +1 For the first -1 day t Load value at time +1 For the first m Heaven and the First m -1 day t The load difference at any given time.

[0025] Normalization is performed using min-max normalization, calculated using the following formula: ;in, The original data, For normalized data, The minimum value of the original data. This represents the maximum value of the original data.

[0026] The minimum and maximum values ​​of the original data were determined by iterating through the historical load data from 2020 to 2021. For example, if the load data range for a certain period is 500 MW to 1500 MW, then 1000 MW is normalized to 0.5. The processing procedure is similar to other normalization methods.

[0027] Feature selection uses the Pearson correlation coefficient to calculate the correlation between each feature and the load data, retaining features with an absolute correlation coefficient of not less than 0.4. The formula for calculating the Pearson correlation coefficient is: ; in, For the first i Each feature data, For the first i Individual load data, The average value of the feature data. This is the average value of the load data. This represents the data sample size. S2: Cluster the preprocessed historical load data, and divide it into at least two load characteristic groups according to different seasons or electricity consumption patterns based on the clustering results.

[0028] In step S2, the DBSCAN clustering algorithm is used to cluster the preprocessed historical load data. In the parameter configuration, the radius eps=410, which can effectively distinguish the load characteristics of different seasons; the minimum number of samples Min_samples=89, which is determined according to the temporal continuity of the load data to ensure the stability of the clustering results.

[0029] The distance is calculated using Euclidean distance, and the formula is as follows: ;in, This represents the Euclidean distance between the i-th sample and the j-th sample; This represents the i-th load characteristic sample; This represents the i-th load characteristic sample; This represents the value of the i-th feature dimension of the i-th sample; This represents the value of the i-th feature dimension of the i-th sample; This represents the total number of feature dimensions for a single load feature sample; This indicates the order of the distance calculation.

[0030] According to the clustering result, at least two load characteristic groups are divided, specifically: in the clustering process, density accessibility analysis is performed on the hourly load data, and two load characteristic groups, i.e., spring-summer season I and autumn-winter season II, are divided, which is consistent with the same cluster clustering result, and each load characteristic group corresponds to different seasons or power consumption modes.

[0031] The silhouette coefficient is evaluated by calculating the average distance of the sample from other samples in the same cluster and the average distance from the nearest sample in the different cluster. The calculation formula of the silhouette coefficient of DBSCAN clustering is:

[0032] wherein, is the average distance of the sample from other samples in the same cluster, is the average distance of the sample from the nearest sample in the different cluster, is the silhouette coefficient; in this embodiment, the silhouette coefficient is 0.4024, which is close to 1, indicating that the clustering effect is good, which meets the evaluation standard of conventional clustering effect. S3: Based on the load characteristic groups, three preliminary prediction results are obtained respectively by using the trained linear regression model, LSTM neural network model and support vector machine model.

[0033] In step S3, the trained linear regression model, LSTM neural network model and support vector machine model are used for each load characteristic group, specifically as follows:

[0034] The linear regression model adopts multivariate linear regression, and the expression is: wherein, is the input feature vector, is the weight parameter, is the bias term, is the first preliminary prediction result of the output.

[0035] Parameter solving: the least squares method is used to minimize the sum of squares of residuals, and the solving formula is: wherein, is the weight parameter estimate value, is the feature matrix, is the transpose of, is the target vector; Training and verification: the training set and the test set are divided by 8:2, and the fitting effect is evaluated by the root mean square error RMSE. This model is suitable for capturing the linear trend of load.

[0036] ​​​​The network architecture of the LSTM neural network model is that the input layer dimension is the number of features, the hidden layer contains 2 LSTM units, the number of units in each layer is 64, and the output layer dimension is 1. The cell state update formula of the LSTM unit is: ; The forget gate output, input gate output, and current time hidden state are calculated by the following corresponding formulas: ; ; ; Among them, is the current time cell state, is the last time cell state, is the forget gate output, is the input gate output, is the candidate cell state, is the current time hidden state, is the last time hidden state, is the current time input, is the weight matrix, is the bias term. The training process uses the Adam optimizer, the learning rate is set to 0.001, the momentum parameters β1=0.9, β2=0.999, the loss function uses mean square error, the iteration number is 50 times, the batch size is 32, and the Dropout layer (proportion 0.2) is used to prevent overfitting. The loss value gradually decreases with the increase of the number of iterations during the training process, and finally stabilizes at about 0.02.

[0037] The support vector machine model uses Support vector regression model, the objective function is: ; wherein, weight vector, relaxation variable, penalty coefficient, sample size; The constraint condition is: ; wherein, is the input feature vector, is the kernel function mapping, is the bias term, is the target value, is the insensitive loss parameter; Kernel function selection: radial basis kernel function is used, and the decision function is: ; wherein, is a Lagrange multiplier, is a kernel function, is a bias term. The penalty coefficient C=10, ε=0.1 in the training parameters, the kernel function parameter γ=0.1 is optimized by grid search, and the model has strong generalization ability when processing high-dimensional features.

[0038] S4: Based on the historical prediction errors of the three preliminary prediction results corresponding to the models, the weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model are determined by using the genetic algorithm, and the combined prediction result is obtained.

[0039] In step S4, the symmetric mean absolute percentage error is used to calculate the historical prediction errors of the three preliminary prediction results corresponding to the models, and the smaller the value, the higher the fitness, which is expressed as: ; wherein, is a predicted value, is an actual value, is a predicted sample size.

[0040] The weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model are determined by using the genetic algorithm, and the specific method is: real number coding is used, each chromosome corresponds to a set of weight coefficients , satisfying + + =1 and ≥0; the fitness function of the genetic algorithm takes the symmetric mean absolute percentage error value of the combined prediction result as the fitness, and the selection operator of the genetic algorithm uses tournament selection, randomly selects 3 individuals from the population each time, and retains the individual with the highest fitness; the crossover operator uses arithmetic crossover, and generates offspring from two parent chromosomes according to the proportion; the mutation operator uses uniform mutation, randomly changes the weight values in the chromosome, the population size is 50, the iteration number is 30 times, and the optimized weight coefficients converge stably, for example, the weights of the spring and summer I type load are 0.2, 0.6 and 0.2.

[0041] S5: The combined prediction result is subjected to inverse normalization processing, and the future load prediction value of the virtual power plant in the preset period (72 hours) is output.

[0042] In step S5, the calculation formula of the inverse normalization processing is: ; wherein, is the output value after inverse normalization processing; is the input value of the inverse normalization processing; is the maximum value of the original load data; is the minimum value of the original load data.

[0043] S6: Model dynamic update every 7 days, when the data drift exceeds the threshold (SMAPE>5%), call the latest actual load data and feature data to retrain the model parameters and optimize the weight coefficient, maintain the stability of the prediction accuracy.

[0044] Adjust the hidden layer unit number of LSTM neural network and the penalty coefficient of support vector machine and other parameters when retraining to ensure that the model adapts to the latest data features, such as increasing the LSTM unit number to 80 during the summer load peak period to improve prediction accuracy. This mechanism refers to the demand for model robustness improvement in similar studies.

[0045] This embodiment is applied to a certain virtual power plant, which contains 2003 10-kilovolt transformers, and needs to realize grid-based prediction of the substation load to provide fine data support for distribution network dispatching. The data processing and model design refer to the typical architecture of substation load prediction.

[0046] Data collection is completed through substation intelligent monitoring terminals, including substation I-class data: User monthly maximum load, industry category (industrial, commercial, residential), geographic topology (latitude and longitude); Substation II-class data: Every 15-minute three-phase current, voltage, active power, and converted to hourly level, the conversion method is to take the average value of 4 15-minute data in each hour; Transformer data: rated capacity, line loss rate, maximum load rate; Other data: air temperature, electricity price, renewable energy quota in green certificate transaction, all data collection frequency is once an hour, consistent with the regular collection frequency of substation data.

[0047] Abnormal value correction in data preprocessing: For the abnormal load value of 200 megawatts at 14:00 on July 1, 2022 in a certain substation, the average load value of 150 megawatts at 13:00, 13:30, 14:30 and 15:00 is calculated by horizontal comparison method. Because 200 megawatts exceeds 150 megawatts by 1.2 times, it is corrected to 150 megawatts.

[0048] Feature selection retains features with absolute correlation coefficient greater than or equal to 0.4 by calculating Pearson correlation coefficient, such as line loss rate (0.45), air temperature (0.52), and industry category (0.43). The selection method refers to the standard of similar feature selection.

[0049] Clustering and grouping use DBSCAN clustering algorithm, and the parameter setting of radius eps=300 adapts to the geographical range of the substation, and the minimum sample number Min_samples=50 is determined based on the spatial distribution characteristics of the substation load data, which is consistent with the similar partitioning idea.

[0050] The clustering process combines the propagation of the substation load and performs density reachability analysis on the preprocessed load data to divide it into three characteristic groups: industrial intensive area, commercial core area, and residential area. The profile coefficient of 0.38 indicates that the clustering result is reasonable.

[0051] In multi-model prediction, the industrial intensive area has long-term trends in load fluctuations, so the LSTM neural network model is mainly used, with the number of hidden layer units adjusted to 80 to capture long-term changes. The commercial core area has significant short-term fluctuations in load, so the support vector machine model is used, with the kernel function parameter γ = 0.08 to adapt to short-term characteristics. The residential area has stable load patterns, so the linear regression model is used.

[0052] Dynamic weighting optimizes the weights through genetic algorithms, with the industrial intensive area weight being 0.1, 0.7, and 0.2, the commercial core area weight being 0.3, 0.2, and 0.5, and the residential area weight being 0.6, 0.2, and 0.2, to ensure that each model complements each other's advantages.

[0053] Grid processing divides the prediction area into 50m x 50m grids based on the substation geographic topology data, calculates the load value of each grid by combining the transformer power supply range and load density, and uses the formula: Grid load = (user proportion in the grid x total predicted load of the substation) x load density coefficient The load density coefficient is corrected according to the substation line loss rate and geographic features, and finally generates a visual grid-based load map, providing spatial fine data for distribution network dispatching.

[0054] Model update mechanism triggers every 15 days, and when the grid load prediction error (SMAPE) exceeds 8%, the clustering model and prediction model are retrained, the LSTM unit number of the industrial intensive area and the SVM parameter of the commercial core area are adjusted, to ensure that the model adapts to the dynamic changes of the substation load and meets the actual needs of the substation fine dispatching.

[0055] Example Two A virtual power plant load prediction system based on multi-model combination is provided in this embodiment, which includes the following modules: The data acquisition and preprocessing module is configured to acquire multi-source heterogeneous data containing virtual power plant historical load data, and preprocess the multi-source heterogeneous data. The clustering module is configured to cluster the preprocessed historical load data, and divide at least two load characteristic groups according to different seasons or power consumption modes according to the clustering result. The preliminary prediction module is configured to obtain three preliminary prediction results respectively by using the trained linear regression model, the LSTM neural network model and the support vector machine model based on the load feature group; The weighting module is configured to determine the weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model by using a genetic algorithm based on the model historical prediction errors corresponding to the three preliminary prediction results, and obtain a combined prediction result. The output module is configured to perform inverse normalization processing on the combined prediction result, and output a load prediction value of the virtual power plant in a future preset time period.

[0056] The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A virtual power plant load forecasting method based on multi-model combination, characterized in that, include: Acquire multi-source heterogeneous data containing historical load data of virtual power plants, and preprocess the multi-source heterogeneous data; Cluster the preprocessed historical load data, and divide it into at least two load characteristic groups according to different seasons or electricity consumption patterns based on the clustering results; Based on the load feature group, three preliminary prediction results were obtained by using the trained linear regression model, LSTM neural network model and support vector machine model respectively. Based on the historical prediction errors of the models corresponding to the three preliminary prediction results, the weight coefficients of the linear regression model, the LSTM neural network model and the support vector machine model are determined by the genetic algorithm, and the combined prediction results are obtained. The combined forecast results are inversely normalized to output the load forecast values ​​for the virtual power plant during the future preset time period.

2. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The multi-source heterogeneous data also includes meteorological data, electricity price data, green certificate trading data, and major event data from virtual power plants; the preprocessing includes timestamp alignment, outlier correction, normalization, and feature filtering, wherein the timestamp alignment is based on the timestamps of historical load data; The outlier correction method employs both horizontal and vertical comparison methods; the normalization process uses min-max normalization. The feature selection process uses Pearson correlation coefficient to calculate the correlation between meteorological data, electricity price data, green certificate trading data, major event data, and historical load data.

3. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The linear regression model adopted is a multiple linear regression model, expressed as: ;in, For the input feature vector, For weight parameters, For bias terms, This is the first preliminary prediction result output.

4. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The cell state update formula for the LSTM unit in the LSTM neural network model is as follows: ； The formulas for the forget gate output, input gate output, and the current hidden state are: ； ； ； in, This represents the current state of the cell. This represents the cell state at the previous moment. Output for the forget gate. For input gate output, Candidate cell state, Hide the current state. This is the hidden state from the previous moment. Enter the current time. This is the weight matrix. This is a bias term.

5. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The support vector machine model adopts The objective function of the support vector regression model is: ;in, Weight vector, As slack variables, The penalty coefficient is... For sample size; Its constraints are: ;in, For the first 1 input feature vector For kernel function mapping, For bias terms, For the target value, This is an insensitive loss parameter.

6. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The historical prediction errors of the model corresponding to the three preliminary prediction results are calculated using the symmetric mean absolute percentage error, and are expressed as follows: ;in, For predicted values, This is the actual value. To predict the sample size.

7. The virtual power plant load forecasting method based on multi-model combination as described in claim 1, characterized in that, The weight coefficients of the linear regression model, LSTM neural network model, and support vector machine model are determined using a genetic algorithm. Specifically, real-number encoding is used, with each chromosome corresponding to a set of weight coefficients. ,satisfy + + =1 and ≥0; The fitness function of the genetic algorithm uses the symmetric average absolute percentage error value of the combined prediction results as the fitness value, and the selection operator of the genetic algorithm adopts tournament selection, the crossover operator adopts arithmetic crossover, and the mutation operator adopts uniform mutation.

8. A virtual power plant load forecasting system based on multi-model combination, characterized in that, Includes the following modules: The data acquisition and preprocessing module is configured to: acquire multi-source heterogeneous data containing historical load data of virtual power plants, and preprocess the multi-source heterogeneous data; The clustering module is configured to: cluster the preprocessed historical load data and divide it into at least two load feature groups according to different seasons or electricity consumption patterns based on the clustering results; The preliminary prediction module is configured to obtain three preliminary prediction results based on the load feature group, using a trained linear regression model, an LSTM neural network model, and a support vector machine model respectively. The weighting module is configured to: determine the weight coefficients of the linear regression model, the LSTM neural network model, and the support vector machine model based on the historical prediction errors of the models corresponding to the three preliminary prediction results, and obtain the combined prediction results; The output module is configured to perform inverse normalization on the combined forecast results and output the load forecast values ​​for the virtual power plant for a future preset period.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the virtual power plant load forecasting method based on multi-model combination as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the virtual power plant load forecasting method based on multi-model combination as described in any one of claims 1-7.

Images