A source-load integrated prediction method based on regression analysis and LSSVM

Abstract

The present invention relates to a source-load integrated prediction method based on regression analysis and LSSVM, comprising the following steps: collecting data; preprocessing the collected data, removing abnormal data, and normalizing the data; calculating the Pearson correlation coefficient r between the power time series of each wind and photovoltaic power station in the forecasted area and the total power series of the wind and photovoltaic power stations in the area, as well as the data accuracy C of each wind and photovoltaic power station; performing STL-based time series decomposition on the time series data of the characteristic wind and photovoltaic power stations in the forecasted area; constructing a regression prediction model; obtaining the final wind power and photovoltaic power forecast for the forecasted area; and obtaining the load forecast result for the forecasted area. The present invention constructs different regression functions for the characteristics of each component through STL time series decomposition to quantify the impact of meteorological factors and time factors on the output of wind farms and photovoltaic power stations. The present invention takes into account the impact of meteorological factors on different components, and the prediction model is more detailed.

Description

Technical Field

[0001] The present invention relates to the technical field of power system source-load prediction, in particular to a source-load integrated prediction method based on regression analysis and LSSVM. Background Art

[0002] Unlike conventional power sources, wind power and photovoltaic power generation have small individual unit capacities, large numbers of units, large output fluctuations, poor predictability, and limited ability to participate in active power regulation. In new power systems characterized by a high proportion of renewable energy and power electronic devices, the source (wind and solar) and loads form a unified entity through the power grid, reducing the controllability of the grid. Accurate and timely forecasting of renewable energy generation power facilitates grid dispatchers to dynamically adjust the output of conventional units, effectively reducing spinning reserve capacity and thus lowering system operating costs. This is the foundation and key to guiding the dispatch and control of new power systems.

[0003] The increasing proportion of wind and solar power connected to the grid, coupled with the increasing demand response load on the load side, is increasing uncertainty in the current power system. Ensuring power quality and ensuring the safe and economic operation of the power system not only places higher demands on the accuracy of wind and solar power forecasts, but also presents new challenges for load forecasting. There is an urgent need to explore new power forecasting methods that can achieve high-precision forecasts of both wind, solar, and load, addressing the dual uncertainties of both the source and the load.

[0004] "Research on a Grid-Connected Photovoltaic Power Generation Forecasting Model Based on KPCA and Hybrid Leapfrog Algorithms" establishes a forecasting model for photovoltaic power plants on rainy days; "Multivariate Regression Load Forecasting Based on Factor and Trend Analysis Feedback" utilizes factor analysis to screen meteorological factors and constructs a multivariate regression model for load under different weather conditions; and "An Ultra-Short-Term Wind Power Forecasting Model Combining CNN and GRU Networks" considers the influence of temporal factors based on meteorological data analysis and establishes a CNN and GRU wind power forecasting model. However, these direct forecasting methods fail to account for the correlations between various meteorological factors, reducing forecast accuracy. Currently, domestic and international scholars have conducted extensive research on wind power, photovoltaic power, and load forecasting, but this research primarily focuses on single-object forecasting. Research on forecasting two or more of these elements—wind, solar, and load—is limited. Furthermore, most of this research fails to consider the correlation between the source (wind, solar, and load) and the load, and instead models and forecasts each element independently.

[0005] Therefore, a more advanced source (wind and solar) and load integrated prediction method with better prediction accuracy is needed. Summary of the Invention

[0006] The purpose of the present invention is to provide a source-load integrated prediction method based on regression analysis and LSSVM, which has a more detailed prediction model and effectively improves the prediction accuracy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a source-load integrated prediction method based on regression analysis and LSSVM, which includes the following steps in sequence:

[0008] (1) Collect power generation data, load data and meteorological data of wind and photovoltaic power stations in the area to be predicted;

[0009] (2) preprocessing the data collected in step (1), removing abnormal data, and performing normalization;

[0010] (3) Calculate the Pearson correlation coefficient r between the power time series of each wind and photovoltaic power station in the forecast area and the total power series of wind and photovoltaic power stations in the area, as well as the data accuracy C of each wind and photovoltaic power station;

[0011] (4) Decompose the time series data of characteristic wind and photovoltaic power stations in the forecast area based on STL;

[0012] (5) Based on the nonlinear multiple regression analysis method, the relationship between the characteristic wind after time series decomposition, the time series of the photovoltaic power station and the meteorological factors of wind speed, temperature, and irradiation intensity is determined, and a regression prediction model is constructed;

[0013] (6) According to the weather forecast, the meteorological factors of the predicted day are substituted into the regression prediction model of the long-term component, periodic fluctuation component and noise component after the decomposition of the characteristic wind and photovoltaic station. After the prediction, the various components are superimposed to obtain the characteristic wind and photovoltaic station power generation prediction power. Based on the characteristic wind and photovoltaic station capacity, the final wind power and photovoltaic power prediction power of the predicted area are obtained;

[0014] (7) The power generation data, meteorological data and load data of the wind and photovoltaic power stations in the area to be predicted are divided into a training set and a validation set. The load forecasting model is obtained by training with the least squares support vector machine (LSSVM) based on the piecewise kernel function. The predicted power generation data and meteorological data of the wind and photovoltaic power stations are substituted into the trained load forecasting model to obtain the load forecast results of the area to be predicted.

[0015] In step (1), the power generation data includes the time-series output power Pw of n wind farms t,i , i=1,2,…,n, and m photovoltaic fields sequential output power Pv t,j , j = 1, 2, ..., m; the load data is the regional power load P Lt , t is the time scale; the meteorological data include radiation intensity f1, ambient wind speed f2, ambient temperature f3, ambient humidity f4 and precipitation f5.

[0016] The step (2) specifically includes the following steps:

[0017] (2a) Eliminate abnormal data:

[0018] f(x)<Q1-1.5×IQR,f(x)> Q3+1.5×IQR (1)

[0019] Where f(x) is the abnormal data, Q1 is the lower quartile of the power generation data and load data of the wind and photovoltaic power stations, Q3 is the upper quartile of the power generation data and load data of the wind and photovoltaic power stations, and IQR is the upper and lower quartile difference of the power generation data and load data of the wind and photovoltaic power stations, that is, Q3-Q1;

[0020] (2b) The power generation data and load data of wind and photovoltaic power stations are unified to the shortest time scale. For missing data on the long time scale, the interpolation method is used to supplement the data, as shown in the following formula:

[0021] (tm)=i=0nL(ti)j=0,j≠intm-tjti-tj (2)

[0022] Among them, L(t i ) is the data at t i The value at the moment, L(t) is the value of the data at the moment t, tm is the moment when the data is missing, ti and tj are the two sampling times near the moment when the data is missing;

[0023] (2c) Perform normalization.

[0024] The step (3) specifically includes the following steps:

[0025] (3a) Calculate the Pearson correlation coefficient r of the power time series of each wind and photovoltaic power station and the total power series of wind and photovoltaic power stations in the region:

[0026]

[0027] Among them, when d=1, it is a wind farm, when d=2, it is a photovoltaic farm, and x t,i,d is the power time series of the i-th wind and photovoltaic power station; y t,d is the total power sequence of wind and photovoltaic power stations in the region; n is the length of the time series; and x t,i,d and y t,d The mean of

[0028] (3b) Calculate the data accuracy C of each wind and solar power station, which is between 0 and 1:

[0029]

[0030] Among them, N p,d,iis the number of abnormal data of the i-th wind and photovoltaic power station in a period of time; N q,d,i is the number of data collected during the same time period for the i-th wind and photovoltaic station;

[0031] (3c) Define RC indicators and select characteristic wind and photovoltaic power stations:

[0032] RC=r+C (5).

[0033] In step (4), the STL-based time series decomposition refers to a time series decomposition method based on robust local weighted regression, and the characteristic wind and photovoltaic power station power time series are decomposed into:

[0034] P d,t =T d,t +C d,t +I d,t (6)

[0035] Among them, when d=1, it is the characteristic wind farm, when d=2, it is the characteristic photovoltaic farm, t represents the time period, P d,t is the characteristic wind and photovoltaic power station power sequence; T d,t is the long-term component; C d,t is the periodic fluctuation component; I d,t is the noise component.

[0036] The step (5) specifically refers to:

[0037] First, based on the nonlinear multiple regression analysis method, the regression equations between each component and time and meteorological factors are obtained:

[0038]

[0039] Among them, I is the number of meteorological indicators, α dT,l , α dc,l , α dI,l is the regression coefficient of each component to meteorological factors, α d1 , β d2 , β d3 is the regression coefficient of each component to time, ε d1 , ε d2 , ε d3 is the regression residual under each component, g d1 (t), g d2 (t), g d3 (t) is the time function corresponding to each component, g d1 (t), g d2 (t), g d3 The formula for (t) is:

[0040]

[0041] Among them, h i (t) is the time function that can fit the long-term component, with linear function being the first choice and exponential function being the second choice; g d2 (t) is the time function of the fitting periodic component, and Fourier is used for fitting; a0, b n 、a n is the combination coefficient in the Fourier function; e i (t) is the time function of the noise component, and the white noise function is used for fitting;

[0042] Combining equations (7) and (8), we can construct regression models of each component of characteristic wind and photovoltaic power station, and then according to equation (6), we can obtain the regression prediction model of characteristic wind and photovoltaic power station;

[0043] The step (6) specifically includes the following steps:

[0044] By calculating the conversion coefficients of characteristic wind power, photovoltaic power station and rated capacity of the region, the total wind power and photovoltaic power station predicted power in the region are obtained;

[0045] (6a) According to the weather forecast, the meteorological factors of the forecast day are substituted into equations (6), (7), and (8) to predict the power generation of a characteristic wind power station p wt,a and b characteristic photovoltaic power station power generation power p vt,b ;

[0046] (6b) Calculate the conversion coefficient between the characteristic wind, the photovoltaic power station and the rated capacity of the region. The conversion coefficient γ is calculated as follows:

[0047]

[0048] Among them, γ w and γ v are wind power and photovoltaic conversion coefficients, P wNj and P vNj are the rated capacities of wind power and photovoltaic power stations in the region, P wNi and P vNi is the rated capacity of the characteristic wind and photovoltaic power stations, and the total wind and photovoltaic power forecast in the region is calculated as:

[0049]

[0050] Among them, P w and P v Forecast power for total wind and photovoltaic power stations in the region.

[0051] The step (7) specifically includes the following steps:

[0052] (7a) Construct the input variable set and output variable set required for load forecasting. The input variable set is x = {f1, f2, f3, f4, f5, Pwt, Pvt} = {x1, x2, x3, x4, x5, x6, x7}, which contains meteorological data and the total power data of wind farms and photovoltaic power stations in the region. The output variable set y = PLt is the regional power load. Then, the set is divided into training set and validation set according to the proportion.

[0053] (7b) Construct a regression function based on LSSVM, whose equation is expressed as y = f(x):

[0054]

[0055] Among them, w is the weight vector, b is the output bias, and φ(x) represents the nonlinear mapping relationship between the input variable and the high-dimensional space;

[0056] (7c) According to the principle of structural risk minimization, in order to solve Equation (9), a slack variable is introduced, and the objective function of the regression problem and the corresponding constraints are:

[0057]

[0058]

[0059] Where c is the penalty function, ξi is the slack variable, and N is the number of samples;

[0060] (7d) Lagrange multipliers are introduced to construct the Lagrange function to solve the above regression problem, as follows:

[0061]

[0062] Where λi is the Lagrange multiplier;

[0063] When the Lagrangian function reaches its extreme value, it has the following formula:

[0064]

[0065] (7e)The solution of formula (13) is:

[0066]

[0067] Among them, λ=[λ1,λ2,…,λn]T, Y=[y1,y2,…,yn]T, Ω={Ωij|i,j=1,2,…,n}, For the selected kernel function, set the piecewise kernel function K(x i ,x j ) is shown in the following formula (15),

[0068]

[0069] Where q is the threshold, σ is the bandwidth, and A is the constant term of the radial basis kernel function;

[0070] (7f) The final load forecasting model is:

[0071]

[0072] Determine whether the load forecasting model meets the accuracy requirements based on the validation set. If the accuracy requirements are not met, return to step (7c) and modify the penalty function, piecewise function threshold, and bandwidth until the accuracy requirements are met. If the accuracy requirements are met, output the load forecasting model according to formula (16).

[0073] (7g) The prediction results are obtained according to the load prediction model:

[0074] According to the trained load forecasting model, the weather forecast and the predicted power of wind and photovoltaic power stations are substituted into formula (16) to obtain the load forecast result.

[0075] It can be seen from the above technical solution that the beneficial effects of the present invention are: first, the present invention decomposes the STL time series and constructs different regression functions for the characteristics of each component to quantify the influence of meteorological factors and time factors on the output of wind farms and photovoltaic power stations. Compared with traditional time series prediction, the present invention takes into account the influence of meteorological factors on different components, and the prediction model is more detailed; second, the present invention takes into account the source-load correlation in the prediction, and establishes a load prediction model based on LSSVM of future wind and solar power generation information. In addition to historical load data and meteorological data, wind and solar power generation power prediction results are also added to the input variables. Compared with a single load prediction, the present invention takes into account the influence of wind and solar output on the load; third, the present invention defines the comprehensive index RC of the power station, and comprehensively selects the characteristic wind and photovoltaic power stations in the region by calculating the correlation of each wind and photovoltaic power station and the data completeness of each wind and photovoltaic power station; fourth, the present invention uses the LSSVM of the piecewise kernel function to process multi-source data, so as to fully utilize the characteristics of the global kernel function and the local kernel function, so as to make the LSSVM training more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0077] like Figure 1 As shown, a source-load integrated prediction method based on regression analysis and LSSVM includes the following steps in sequence:

[0078] (1) Collect power generation data, load data and meteorological data of wind and photovoltaic power stations in the area to be predicted;

[0079] (2) preprocessing the data collected in step (1), removing abnormal data, and performing normalization;

[0080] (3) Calculate the Pearson correlation coefficient r between the power time series of each wind and photovoltaic power station in the forecast area and the total power series of wind and photovoltaic power stations in the area, as well as the data accuracy C of each wind and photovoltaic power station;

[0081] (4) Decompose the time series data of characteristic wind and photovoltaic power stations in the forecast area based on STL;

[0082] (5) Based on the nonlinear multiple regression analysis method, the relationship between the characteristic wind after time series decomposition, the time series of the photovoltaic power station and the meteorological factors of wind speed, temperature, and irradiation intensity is determined, and a regression prediction model is constructed;

[0083] (6) According to the weather forecast, the meteorological factors of the predicted day are substituted into the regression prediction model of the long-term component, periodic fluctuation component and noise component after the decomposition of the characteristic wind and photovoltaic station. After the prediction, the various components are superimposed to obtain the characteristic wind and photovoltaic station power generation prediction power. Based on the characteristic wind and photovoltaic station capacity, the final wind power and photovoltaic power prediction power of the predicted area are obtained;

[0084] (7) The power generation data, meteorological data and load data of the wind and photovoltaic power stations in the area to be predicted are divided into a training set and a validation set. The load forecasting model is obtained by training with the least squares support vector machine (LSSVM) based on the piecewise kernel function. The predicted power generation data and meteorological data of the wind and photovoltaic power stations are substituted into the trained load forecasting model to obtain the load forecast results of the area to be predicted.

[0085] In step (1), the power generation data includes the time-series output power Pw of n wind farms t,i , i=1,2,…,n, and m photovoltaic fields sequential output power Pv t,j , j = 1, 2, ..., m; the load data is the regional power load P Lt , t is the time scale; the meteorological data include radiation intensity f1, ambient wind speed f2, ambient temperature f3, ambient humidity f4 and precipitation f5.

[0086] The step (2) specifically includes the following steps:

[0087] (2a) Eliminate abnormal data:

[0088] f(x)<Q1-1.5×IQR,f(x)> Q3+1.5×IQR (1)

[0089] Where f(x) is the abnormal data, Q1 is the lower quartile of the power generation data and load data of the wind and photovoltaic power stations, Q3 is the upper quartile of the power generation data and load data of the wind and photovoltaic power stations, and IQR is the upper and lower quartile difference of the power generation data and load data of the wind and photovoltaic power stations, that is, Q3-Q1;

[0090] (2b) The power generation data and load data of wind and photovoltaic power stations are unified to the shortest time scale. For missing data on the long time scale, the interpolation method is used to supplement the data, as shown in the following formula:

[0091] (tm)=i=0nL(ti)j=0,j≠intm-tjti-tj (2)

[0092] Among them, L(t i ) is the data at t i The value at the moment, L(t) is the value of the data at the moment t, tm is the moment when the data is missing, ti and tj are the two sampling times near the moment when the data is missing;

[0093] (2c) Perform normalization.

[0094] The step (3) specifically includes the following steps:

[0095] (3a) Calculate the Pearson correlation coefficient r of the power time series of each wind and photovoltaic power station and the total power series of wind and photovoltaic power stations in the region:

[0096]

[0097] Among them, when d=1, it is a wind farm, when d=2, it is a photovoltaic farm, and x t,i,d is the power time series of the i-th wind and photovoltaic power station; y t,d is the total power sequence of wind and photovoltaic power stations in the region; n is the length of the time series; and x t,i,d and y t,d The value of r ranges from -1 to 1. The larger the Pearson correlation coefficient, the stronger the correlation.

[0098] (3b) Calculate the data accuracy C of each wind and solar power station, which is between 0 and 1:

[0099]

[0100] Among them, N p,d,i is the number of abnormal data of the i-th wind and photovoltaic power station in a period of time; N q,d,i is the number of data collected during the same time period for the i-th wind and photovoltaic station;

[0101] (3c) Define RC indicators and select characteristic wind and photovoltaic power stations:

[0102] RC=r+C (5)

[0103] The closer the RC index is to 2, the more the selected wind and solar power stations can represent the wind and solar output characteristics of the entire region.

[0104] In step (4), the STL-based time series decomposition refers to a time series decomposition method based on robust local weighted regression, and the characteristic wind and photovoltaic power station power time series are decomposed into:

[0105] P d,t =T d,t +C d,t +I d,t (6)

[0106] Among them, when d=1, it is the characteristic wind farm, when d=2, it is the characteristic photovoltaic farm, t represents the time period, P d,t is the characteristic wind and photovoltaic power station power sequence; T d,t is the long-term component; C d,t is the periodic fluctuation component; I d,t is the noise component.

[0107] The STL decomposition method refers to a time series decomposition method based on robust locally weighted regression, that is, a seasonal trend decomposition procedure based on locally weighted scatterplotsmoothing, which can decompose a time series according to any period.

[0108] Considering that the random fluctuation of meteorological factors has a great impact on wind and solar power generation, and there are complex correlations between meteorological factors, if the original meteorological data is continued to be used to establish the prediction model, some information will be covered, affecting the accuracy of the prediction model. Therefore, a multivariate nonlinear regression analysis is performed on the decomposed characteristic wind and photovoltaic power station components.

[0109] The step (5) specifically refers to:

[0110] First, based on the nonlinear multiple regression analysis method, the regression equations between each component and time and meteorological factors are obtained:

[0111]

[0112] Among them, I is the number of meteorological indicators, α dT,l , α dc,l , α dI,l is the regression coefficient of each component to meteorological factors, α d1, β d2 , β d3 is the regression coefficient of each component to time, ε d1 , ε d2 , ε d3 is the regression residual under each component, g d1 (t), g d2 (t), g d3 (t) is the time function corresponding to each component, g d1 (t), g d2 (t), g d3 The formula for (t) is:

[0113]

[0114] Among them, h i (t) is the time function that can fit the long-term component, with linear function being the first choice and exponential function being the second choice; g d2 (t) is the time function of the fitting periodic component, and Fourier is used for fitting; a0, b n 、a n is the combination coefficient in the Fourier function; e i (t) is the time function of the noise component, and the white noise function is used for fitting;

[0115] Combining equations (7) and (8), we can construct regression models of each component of characteristic wind and photovoltaic power station, and then according to equation (6), we can obtain the regression prediction model of characteristic wind and photovoltaic power station;

[0116] The step (6) specifically includes the following steps:

[0117] By calculating the conversion coefficients of characteristic wind power, photovoltaic power station and rated capacity of the region, the total wind power and photovoltaic power station predicted power in the region are obtained;

[0118] (6a) According to the weather forecast, the meteorological factors of the forecast day are substituted into equations (6), (7), and (8) to predict the power generation of a characteristic wind power station p wt,a and b characteristic photovoltaic power station power generation power p vt,b ;

[0119] (6b) Calculate the conversion coefficient between the characteristic wind, the photovoltaic power station and the rated capacity of the region. The conversion coefficient γ is calculated as follows:

[0120]

[0121] Among them, γ w and γ v are wind power and photovoltaic conversion coefficients, P wNj and P vNj are the rated capacities of wind power and photovoltaic power stations in the region, PwNi and P vNi is the rated capacity of the characteristic wind and photovoltaic power stations, and the total wind and photovoltaic power forecast in the region is calculated as:

[0122]

[0123] Among them, P w and P v Forecast power for total wind and photovoltaic power stations in the region.

[0124] To account for the interactive coupling between power sources and loads, a load forecasting model based on the LSSVM with a piecewise kernel function was established that incorporates wind and solar power generation information. This model uses historical load data as well as the historical power of wind and solar power plants as input variables. LSSVM is a machine learning method used to solve nonlinear regression problems. It is characterized by a small sample size and high regression accuracy. However, LSSVM typically uses a radial basis kernel function, which can lead to insufficient accuracy when processing multi-source data. Therefore, a piecewise kernel function approach is used to fully utilize both local and global kernel functions to improve forecasting accuracy. A load forecasting model based on the LSSVM with a piecewise kernel function was established. The LSSVM mentioned above refers to a least squares support vector machine (LSSVM).

[0125] The step (7) specifically includes the following steps:

[0126] (7a) Construct the input variable set and output variable set required for load forecasting. The input variable set is x = {f1, f2, f3, f4, f5, Pwt, Pvt} = {x1, x2, x3, x4, x5, x6, x7}, which contains meteorological data and the total power data of wind farms and photovoltaic power stations in the region. The output variable set y = PLt is the regional power load. Then, the set is divided into training set and validation set according to the proportion.

[0127] (7b) Construct a regression function based on LSSVM, whose equation is expressed as y = f(x):

[0128]

[0129] Among them, w is the weight vector, b is the output bias, and φ(x) represents the nonlinear mapping relationship between the input variable and the high-dimensional space;

[0130] (7c) According to the principle of structural risk minimization, in order to solve Equation (9), a slack variable is introduced, and the objective function of the regression problem and the corresponding constraints are:

[0131]

[0132]

[0133] Where c is the penalty function, ξi is the slack variable, and N is the number of samples;

[0134] (7d) Lagrange multipliers are introduced to construct the Lagrange function to solve the above regression problem, as follows:

[0135]

[0136] Among them, λ i is the Lagrange multiplier;

[0137] When the Lagrangian function reaches its extreme value, it has the following formula:

[0138]

[0139] (7e)The solution of formula (13) is:

[0140]

[0141] Among them, λ=[λ1,λ2,…,λn]T, Y=[y1,y2,…,yn]T, Ω={Ωij|i,j=1,2,…,n}, For the selected kernel function, set the piecewise kernel function K(x i ,x j ) is shown in the following formula (15),

[0142]

[0143] Where q is the threshold, σ is the bandwidth, and A is the constant term of the radial basis kernel function;

[0144] (7f) The final load forecasting model is:

[0145]

[0146] Determine whether the load forecasting model meets the accuracy requirements based on the validation set. If the accuracy requirements are not met, return to step (7c) and modify the penalty function, piecewise function threshold, and bandwidth until the accuracy requirements are met. If the accuracy requirements are met, output the load forecasting model according to formula (16).

[0147] (7g) The prediction results are obtained according to the load prediction model:

[0148] According to the trained load forecasting model, the weather forecast and the predicted power of wind and photovoltaic power stations are substituted into formula (16) to obtain the load forecast result.

[0149] In summary, the present invention decomposes the STL time series and constructs different regression functions for the characteristics of each component to quantify the impact of meteorological factors and time factors on the output of wind farms and photovoltaic power stations. Compared with traditional time series prediction, the present invention takes into account the impact of meteorological factors on different components, and the prediction model is more detailed; the present invention takes into account the source-load correlation in the prediction, and establishes a load prediction model based on LSSVM for future wind and solar power generation information. In addition to historical load data and meteorological data, wind and solar power generation power prediction results are also added to the input variables. Compared with a single load prediction, the present invention takes into account the impact of wind and solar output on the load.

Claims

1. A source-load integrated prediction method based on regression analysis and LSSVM, characterized by: The method comprises the following steps in sequence: (1) Collect power generation data, load data and meteorological data of wind and photovoltaic power stations in the area to be predicted; (2) preprocessing the data collected in step (1), removing abnormal data, and performing normalization; (3) Calculate the Pearson correlation coefficient r between the power time series of each wind and photovoltaic power station in the forecast area and the total power series of wind and photovoltaic power stations in the area, as well as the data accuracy C of each wind and photovoltaic power station; (4) Decompose the time series data of characteristic wind and photovoltaic power stations in the forecast area based on STL; (5) Based on the nonlinear multiple regression analysis method, the relationship between the characteristic wind after time series decomposition, the time series of the photovoltaic power station and the meteorological factors of wind speed, temperature, and irradiation intensity is determined, and a regression prediction model is constructed; (6) According to the weather forecast, the meteorological factors of the predicted day are substituted into the regression prediction model of the long-term component, periodic fluctuation component and noise component after the decomposition of the characteristic wind and photovoltaic station. After the prediction, the various components are superimposed to obtain the characteristic wind and photovoltaic station power generation prediction power. Based on the characteristic wind and photovoltaic station capacity, the final wind power and photovoltaic power prediction power of the predicted area are obtained; (7) The power generation data, meteorological data, and load data of the wind and photovoltaic power stations in the area to be predicted are divided into a training set and a validation set. The load forecasting model is obtained by training with the least squares support vector machine (LSSVM) based on the piecewise kernel function. The predicted power generation data and meteorological data of the wind and photovoltaic power stations are substituted into the trained load forecasting model to obtain the load forecast results for the area to be predicted. The step (7) specifically includes the following steps: (7a) Construct the input variable set and output variable set required for load forecasting. The input variable set is x = {f1, f2, f3, f4, f5, Pwt, Pvt} = {x1, x2, x3, x4, x5, x6, x7}, which contains meteorological data and the total power data of wind farms and photovoltaic power stations in the region. The output variable set y = PLt is the regional power load. Then, the set is divided into training set and validation set according to the proportion. (7b) Construct a regression function based on LSSVM, whose equation is expressed as y = f(x): Among them, w is the weight vector, b is the output bias, and φ(x) represents the nonlinear mapping relationship between the input variable and the high-dimensional space; (7c) According to the principle of structural risk minimization, in order to solve Equation (9), a slack variable is introduced, and the objective function of the regression problem and the corresponding constraints are: Where c is the penalty function, ξi is the slack variable, and N is the number of samples; (7d) Lagrange multipliers are introduced to construct the Lagrange function to solve the above regression problem, as follows: Among them, λ i is the Lagrange multiplier; When the Lagrangian function reaches its extreme value, it has the following formula: (7e)The solution of formula (13) is: Among them, λ=[λ1,λ2,…,λn]T, Y=[y1,y2,…,yn]T, Ω={Ωij|i,j=1,2,…,n}, For the selected kernel function, set the piecewise kernel function K(x i ,x j ) is shown in the following formula (15), Where q is the threshold, σ is the bandwidth, and A is the constant term of the radial basis kernel function; (7f) The final load forecasting model is:

2. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: In step (1), the power generation data includes the time-series output power Pw of n wind farms t,i , i=1,2,…,n, and m photovoltaic fields sequential output power Pv t,j , j = 1, 2, ..., m; the load data is the regional power load P Lt , t is the time scale; the meteorological data include radiation intensity f1, ambient wind speed f2, ambient temperature f3, ambient humidity f4 and precipitation f5.

3. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: The step (2) specifically includes the following steps: (2a) Eliminate abnormal data: f(x)<Q1-1.5×IQR,f(x)> Q3+1.5×IQR (1) Where f(x) is abnormal data, Q1 is the lower quartile of the power generation data and load data of the wind and photovoltaic power stations, Q3 is the upper quartile of the power generation data and load data of the wind and photovoltaic power stations, and IQR is the upper and lower quartile difference of the power generation data and load data of the wind and photovoltaic power stations, that is, Q3-Q1; (2b) The power generation data and load data of wind and photovoltaic power stations are unified to the shortest time scale. For missing data on the long time scale, the interpolation method is used to supplement the data, as shown in the following formula: (tm)=i=0nL(ti)j=0,j≠intm-tjti-tj (2) Among them, L(t i ) is the data at t i The value at the moment, L(t) is the value of the data at the moment t, tm is the moment when the data is missing, ti and tj are the two sampling times near the moment when the data is missing; (2c) Perform normalization.

4. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: The step (3) specifically includes the following steps: (3a) Calculate the Pearson correlation coefficient r of the power time series of each wind and photovoltaic power station and the total power series of wind and photovoltaic power stations in the region: Among them, when d=1, it is a wind farm, when d=2, it is a photovoltaic farm, and x t,i,d is the power time series of the i-th wind and photovoltaic power station; y t,d is the total power sequence of wind and photovoltaic power stations in the region; n is the length of the time series; and x t,i,d and y t,d The mean of (3b) Calculate the data accuracy C of each wind and solar power station, which is between 0 and 1: Among them, N p,d,i is the number of abnormal data of the i-th wind and photovoltaic power station in a period of time; N q,d,i is the number of data collected during the same time period for the i-th wind and photovoltaic station; (3c) Define RC indicators and select characteristic wind and photovoltaic power stations: RC=r+C (5).

5. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: In step (4), the STL-based time series decomposition refers to a time series decomposition method based on robust local weighted regression, and the characteristic wind and photovoltaic power station power time series are decomposed into: P d,t =T d,t +C d,t +I d,t (6) Among them, when d=1, it is the characteristic wind farm, when d=2, it is the characteristic photovoltaic farm, t represents the time period, P d,t is the characteristic wind and photovoltaic power station power sequence; T d,t is the long-term component; C d,t is the periodic fluctuation component; I d,t is the noise component.

6. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: The step (5) specifically refers to: First, based on the nonlinear multiple regression analysis method, the regression equations between each component and time and meteorological factors are obtained: Among them, I is the number of meteorological indicators, α dT,l , α dc,l , α dI,l is the regression coefficient of each component to meteorological factors, β d1 , β d2 , β d3 is the regression coefficient of each component to time, ε d1 , ε d2 , ε d3 is the regression residual under each component, g d1 (t), g d2 (t), g d3 (t) is the time function corresponding to each component, g d1 (t), g d2 (t), g d3 The formula for (t) is: Among them, h i (t) is the time function that can fit the long-term component, with linear function being the first choice and exponential function being the second choice; g d2 (t) is the time function of the fitting periodic component, and Fourier is used for fitting; a0, b n 、a n is the combination coefficient in the Fourier function; e i (t) is the time function of the noise component, and the white noise function is used for fitting; Combining equations (7) and (8), we can construct regression models of each component of characteristic wind and photovoltaic power station, and then according to equation (6), we can obtain the regression prediction model of characteristic wind and photovoltaic power station; The step (6) specifically includes the following steps: By calculating the conversion coefficients of characteristic wind power, photovoltaic power station and rated capacity of the region, the total wind power and photovoltaic power station predicted power in the region are obtained; (6a) According to the weather forecast, the meteorological factors of the forecast day are substituted into equations (6), (7), and (8) to predict the power generation of a characteristic wind power station p wt,a and b characteristic photovoltaic power station power generation power p vt,b ; (6b) Calculate the conversion coefficient between the characteristic wind, the photovoltaic power station and the rated capacity of the region. The conversion coefficient γ is calculated as follows: Among them, γ w and γ v are wind power and photovoltaic conversion coefficients, P wNj and P vNj are the rated capacities of wind power and photovoltaic power stations in the region, P wNi and P vNi is the rated capacity of the characteristic wind and photovoltaic power stations, and the total wind and photovoltaic power forecast in the region is calculated as: Among them, P w and P v Forecast power for total wind and photovoltaic power stations in the region.

7. The source-load integrated prediction method based on regression analysis and LSSVM according to claim 1 is characterized in that: Determine whether the load forecasting model meets the accuracy requirements based on the validation set. If the accuracy requirements are not met, return to step (7c) and modify the penalty function, piecewise function threshold, and bandwidth until the accuracy requirements are met. If the accuracy requirements are met, output the load forecasting model according to formula (16). (7g) The prediction results are obtained according to the load prediction model: According to the trained load forecasting model, the weather forecast and the predicted power of wind and photovoltaic power stations are substituted into formula (16) to obtain the load forecast result.

Images