1. A short-term power load prediction method based on SFO-TSVR is characterized by comprising the following steps:

s1, acquiring historical load data, meteorological data and corresponding date type data as original sample data;

s2, dividing original sample data into a training set and a test set;

s3, setting an input data sequence and constructing a TSVR model;

s4, training the TSVR model based on the training set and by adopting an SFO algorithm;

s5, verifying the trained TSVR model based on the test set, if the verification is passed, executing the step S6, otherwise, returning to the step S3;

and S6, inputting the relevant data of the actual day to be predicted into the trained TSVR model according to the set input data sequence to obtain the predicted value of the power load of the actual day to be predicted.

2. The SFO-TSVR-based short-term power load forecasting method of claim 1, wherein the step S1 is specifically to normalize the historical load data and the meteorological data and quantize the corresponding date type data to be in the range of [0,1] so as to obtain the original sample data.

3. The SFO-TSVR-based short-term power load forecasting method of claim 1, wherein the step S3 comprises the following steps:

s31, setting the input data sequence as: the weather data, the load data and the date type of the day to be predicted, and the weather data and the date type of the day to be predicted;

s32, establishing a regression function of the TSVR model:

f(x)＝(f₁(x)+f₁(x))/2

wherein f (x) is a regression function of the TSVR model, f₁(x) And f₂(x) Upper and lower bound decision functions, w, of the TSVR model, respectively₁And w₂Are respectively decision functions f₁(x) And f₂(x) Weight coefficient of (b)₁And b₂Are respectively decision functions f₁(x) And f₂(x) The bias term of (1).

4. The SFO-TSVR-based short-term power load forecasting method of claim 3, wherein the weighting factors and the bias terms in the step S32 are specifically:

G＝[X e]

f＝Y-eε₁

h＝Y+eε₂

X＝[x₁,x₂,...,x_n]^T

Y＝[y₁,y₂,...,y_n]^T

where α and β are Lagrangian operators, X is the input matrix, Y is the output matrix, ε₁、ε₂To insensitive loss factor, ξ₁And xi₂For the relaxation variable, e is a column vector with elements all 1.

5. The SFO-TSVR-based short-term power load forecasting method of claim 4, wherein the step S4 specifically comprises the steps of:

s41, determining the penalty parameter C of the TSVR to be optimized₁、C₂Insensitive loss factor ε₁And ε₂And a Gaussian kernel parameter sigma, and setting a group of parameters of the TSVR corresponding to the position of each flag fish in the SFO algorithm;

s42, determining the fitness function of the SFO:

wherein M is the number of training set samples,and y_kRespectively representing the actual output value and the expected output value of the training sample;

s44, randomly initializing the positions of the flag fish population and the sardine population according to the feasible region of the parameter to be optimized, and setting the relevant parameters of the SFO algorithm;

s44, calculating initial fitness values of the flag fish population and the sardine population;

s45, determining the optimal individual of the fitness value of the flag fish population and the optimal individual of the fitness value of the sardine population in the initial state, and respectively recording the optimal individual as the flag elegans and the optimal individual as injured sardines;

s46, updating the self position of the flag fish population according to the position of the Elfin gobius elegans and the position of the injured sardine;

s47, the sardine population updates the position according to the attacking force of the flag fish and the orientation of the Elfin flag fish;

s48, recalculating the fitness values of the flag fishes and the sardines and comparing the fitness values, wherein if the fitness values of the sardines are better than the flag fishes, the flag fishes occupy the positions of the sardines and remove the sardines from the population, and the method can be specifically expressed as follows:

wherein the content of the first and second substances,is the position of the flag fish at present,as the current sardine position, S_iIs the fitness of sardines to, SF_iThe fitness value of the flag fish is obtained;

and S49, judging whether the maximum iteration number is reached, if so, outputting the position of the current Ellissima fish, namely the optimal parameter of the TSVR, and otherwise, returning to the step S46.

6. The SFO-TSVR-based short-term power load forecasting method of claim 5, wherein the flag fish population update position in step S46 is calculated by the following formula:

wherein the content of the first and second substances,andrespectively representing the positions of the current Elaphe arguin and the injured sardine, wherein r is a random number uniformly distributed between 0 and 1, and lambda is_iIs a dynamically changing coefficient, whose expression is:

wherein N is_SFAnd N_SRepresenting the number of flag fish and sardine, respectively.

7. The SFO-TSVR-based short-term power load forecasting method of claim 6, wherein the calculation formula of the updated sardine population position in the step S47 is as follows:

wherein the content of the first and second substances,the current position of the sardine is shown, the AP shows the attack force of the flag fish, and the attack force is linearly reduced after each attack, and the calculation formula is as follows:

AP＝A*(1-2Itr*ε)

wherein A and epsilon are the attack force change coefficient of the flag fish, and Itr is the current iteration number.

8. The SFO-TSVR-based short-term power load forecasting method of claim 1, wherein the step S5 comprises the following steps:

s51, selecting corresponding data from the test set to input the trained TSVR model according to the set input data sequence to obtain a corresponding verification predicted value;

s52, carrying out reverse normalization processing on the verification prediction value to obtain a verification prediction result;

and S53, evaluating the verification prediction result based on the corresponding real load data in the test set, if the evaluation value is less than or equal to a preset threshold value, passing the verification, and executing the step S6, otherwise, returning to the step S4.

9. The SFO-TSVR-based short-term power load forecasting method of claim 8, wherein the step S53 is implemented by evaluating the verification forecast result by calculating a mean absolute percentage error.

10. The SFO-TSVR based short term power load forecasting method of claim 9, wherein the average absolute percentage error is calculated as:

wherein MAPE is the mean absolute percentage error, K is the number of sample points to be predicted,and yi are the verified predicted value and the true value of the load data, respectively.

Background

The short-term load prediction is a basic component of the economic dispatching of the power system and is also an important guarantee for the safe operation of the power system. With the rapid development of social economy, the demand and daily increase of electric energy of various industries, meanwhile, the demand on the power load prediction precision is gradually improved, the power production and dispatching operation plan can be reasonably arranged by improving the load prediction precision, and the economic benefit of a power grid is also greatly influenced.

With the development of the fields of computer technology, artificial intelligence and the like, the current short-term power load prediction methods mainly comprise methods of support vector regression, random forest, deep neural network and the like. The deep neural network has the advantages of strong learning ability, good feature extraction effect and the like, but has higher requirements on data scale and hardware; the support vector regression has the advantages of excellence in processing small sample data, strong generalization capability and the like, but has the defects of low running speed and relatively low precision; in addition, Twin Support Vector Regression (TSVR) is based on TSVR, a complex convex optimization problem is converted into two simple convex quadratic programming problems by using two non-parallel hyperplanes, thereby greatly improving the training efficiency and fitting ability of the model, however, twin support vector regression is a parameter-dependent model, the selection of parameters directly affects the prediction accuracy of the model, and the manual parameter adjustment manner is easily affected by subjective experience, and the traditional intelligent optimization algorithm also has certain limitations in the aspects of optimization ability and convergence speed. The above all affect the accuracy and efficiency of the prediction result.

Disclosure of Invention

The invention aims to overcome the defects in the prior art and provide a short-term power load prediction method based on SFO (sailfish optimization) -TSVR (short-term power load forecasting) so as to achieve the purpose of efficient and accurate prediction.

The purpose of the invention can be realized by the following technical scheme: a short-term power load prediction method based on SFO-TSVR comprises the following steps:

s1, acquiring historical load data, meteorological data and corresponding date type data as original sample data;

s2, dividing original sample data into a training set and a test set;

s3, setting an input data sequence and constructing a TSVR model;

s4, training the TSVR model based on the training set and by adopting an SFO algorithm;

s5, verifying the trained TSVR model based on the test set, if the verification is passed, executing the step S6, otherwise, returning to the step S3;

and S6, inputting the relevant data of the actual day to be predicted into the trained TSVR model according to the set input data sequence to obtain the predicted value of the power load of the actual day to be predicted.

Further, step S1 is specifically to perform normalization processing on the historical load data and the meteorological data, and quantize the corresponding date type data to the range of [0,1], so as to obtain the original sample data.

Further, the step S3 specifically includes the following steps:

s31, setting the input data sequence as: the weather data, the load data and the date type of the day to be predicted, and the weather data and the date type of the day to be predicted;

s32, establishing a regression function of the TSVR model:

f(x)＝(f₁(x)+f₁(x))/2

wherein f (x) is a regression function of the TSVR model, f₁(x) And f₂(x) Upper and lower bound decision functions, w, of the TSVR model, respectively₁And w₂Are respectively decision functions f₁(x) And f₂(x) Weight coefficient of (b)₁And b₂Are respectively decision functions f₁(x) And f₂(x) The bias term of (1).

Further, the weighting factor and the bias term in step S32 are specifically:

G＝[X e]

f＝Y-eε₁

h＝Y+eε₂

X＝[x₁,x₂,…,x_n]^T

Y＝[y₁,y₂,…,y_n]^T

where α and β are Lagrangian operators, X is the input matrix, Y is the output matrix, ε₁、ε₂To insensitive loss factor, ξ₁And xi₂For the relaxation variable, e is a column vector with elements all 1.

Further, the step S4 specifically includes the following steps:

s41, determining the penalty parameter C of the TSVR to be optimized₁、C₂Insensitive loss factor ε₁And ε₂And a Gaussian kernel parameter sigma, and setting a group of parameters of the TSVR corresponding to the position of each flag fish in the SFO algorithm;

s42, determining the fitness function of the SFO:

wherein M is the number of training set samples,and y_kRespectively representing the actual output value and the expected output value of the training sample;

s44, randomly initializing the positions of the flag fish population and the sardine population according to the feasible region of the parameter to be optimized, and setting the relevant parameters of the SFO algorithm;

s44, calculating initial fitness values of the flag fish population and the sardine population;

s45, determining the optimal individual of the fitness value of the flag fish population and the optimal individual of the fitness value of the sardine population in the initial state, and respectively recording the optimal individual as the flag elegans and the optimal individual as injured sardines;

s46, updating the self position of the flag fish population according to the position of the Elfin gobius elegans and the position of the injured sardine;

s47, the sardine population updates the position according to the attacking force of the flag fish and the orientation of the Elfin flag fish;

s48, recalculating the fitness values of the flag fishes and the sardines and comparing the fitness values, wherein if the fitness values of the sardines are better than the flag fishes, the flag fishes occupy the positions of the sardines and remove the sardines from the population, and the method can be specifically expressed as follows:

if f(S_i)<f(SF_i)

wherein the content of the first and second substances,is the position of the flag fish at present,as the current sardine position, S_iIs the fitness of sardines to, SF_iThe fitness value of the flag fish is obtained;

and S49, judging whether the maximum iteration number is reached, if so, outputting the position of the current Ellissima fish, namely the optimal parameter of the TSVR, and otherwise, returning to the step S46.

Further, the calculation formula of the flag fish population update position in step S46 is as follows:

wherein the content of the first and second substances,andrespectively representing the positions of the current Elaphe arguin and the injured sardine, wherein r is a random number uniformly distributed between 0 and 1, and lambda is_iIs a dynamically changing coefficient, whose expression is:

wherein N is_SFAnd N_SRepresenting the number of flag fish and sardine, respectively.

Further, the calculation formula of the updated sardine population position in step S47 is as follows:

wherein the content of the first and second substances,the current position of the sardine is shown, the AP shows the attack force of the flag fish, and the attack force is linearly reduced after each attack, and the calculation formula is as follows:

AP＝A*(1-2Itr*ε)

wherein A and epsilon are the attack force change coefficient of the flag fish, and Itr is the current iteration number.

Further, the step S5 specifically includes the following steps:

s51, selecting corresponding data from the test set to input the trained TSVR model according to the set input data sequence to obtain a corresponding verification predicted value;

s52, carrying out reverse normalization processing on the verification prediction value to obtain a verification prediction result;

and S53, evaluating the verification prediction result based on the corresponding real load data in the test set, if the evaluation value is less than or equal to a preset threshold value, passing the verification, and executing the step S6, otherwise, returning to the step S4.

Further, the step S53 specifically evaluates the verification prediction result by calculating a mean absolute percentage error.

Further, the calculation formula of the average absolute percentage error is as follows:

wherein MAPE is the mean absolute percentage error, K is the number of sample points to be predicted,and y_iRespectively, the verification predicted value and the true value of the load data.

Compared with the prior art, the method has the advantages that the SFO-TSVR short-term power load prediction model is constructed, the parameters of the TSVR model are automatically adjusted by adopting the SFO algorithm, the efficiency is higher and more stable, the convergence speed of the TSVR model training can be accelerated, the optimizing precision of the TSVR model training is ensured, and the accuracy and the prediction efficiency of the output prediction result of the TSVR model after training are reliably ensured;

the short-term power load prediction model of the SFO-TSVR, which is established by the invention, is established on the basis of empirical risk minimization, so that the method can be suitable for solving large-scale data, overcomes the defect that support vector regression is not good at processing a large amount of data, and has better performance on training small sample data; the SFO-TSVR model created by the method has higher prediction precision relative to models such as support vector regression, long-term and short-term memory network and twin support vector regression.

Drawings

FIG. 1 is a schematic flow diagram of the process of the present invention;

FIG. 2 is a schematic diagram of an exemplary load prediction process;

FIG. 3 is a diagram illustrating a comparison between a predicted load value and an actual load value in an embodiment;

FIG. 4 is a schematic diagram illustrating comparison of prediction results between the method of the present invention and a conventional prediction method in an embodiment;

FIG. 5 is a schematic diagram illustrating comparison of model evaluation indexes between the method of the present invention and a conventional prediction method in the embodiment.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.

Examples

As shown in FIG. 1, a short-term power load prediction method based on SFO-TSVR comprises the following steps:

s1, acquiring historical load data, meteorological data and corresponding date type data to be used as original sample data, specifically, performing normalization processing on the historical load data and the meteorological data, and quantizing the corresponding date type data to be in a [0,1] range to obtain the original sample data;

s2, dividing original sample data into a training set and a test set, wherein the data of 87.5% of the sample data set can be used as the training set, and the data of 12.5% of the sample data set can be used as the test set;

s3, setting an input data sequence, and constructing a TSVR model, specifically:

s31, setting the input data sequence as: the weather data, the load data and the date type of the day to be predicted, and the weather data and the date type of the day to be predicted;

s32, establishing a regression function of the TSVR model:

f(x)＝(f₁(x)+f₁(x))/2

G＝[X e]

f＝Y-eε₁

h＝Y+eε₂

X＝[x₁,x₂,…,x_n]^T

Y＝[y₁,y₂,…,y_n]^T

wherein f (x) is a regression function of the TSVR model, f₁(x) And f₂(x) Upper and lower bound decision functions, w, of the TSVR model, respectively₁And w₂Are respectively decision functions f₁(x) And f₂(x) Weight coefficient of (b)₁And b₂Are respectively decision functions f₁(x) And f₂(x) Alpha and beta are Lagrangian operators, X is the input matrix, Y is the output matrix, epsilon₁、ε₂To insensitive loss factor, ξ₁And xi₂Is a relaxation variable, e is a column vector whose elements are all 1;

s4, training the TSVR model based on the training set and by adopting an SFO algorithm, specifically:

s41, determining the penalty parameter C of the TSVR to be optimized₁、C₂Insensitive loss factor ε₁And ε₂And a Gaussian kernel parameter sigma, and setting a group of parameters of the TSVR corresponding to the position of each flag fish in the SFO algorithm;

s42, determining the fitness function of the SFO:

wherein M is the number of training set samples,and y_kRespectively representing the actual output value and the expected output value of the training sample;

s44, randomly initializing the positions of the flag fish population and the sardine population according to the feasible region of the parameter to be optimized, and setting the relevant parameters of the SFO algorithm;

s44, calculating initial fitness values of the flag fish population and the sardine population;

s45, determining the optimal individual of the fitness value of the flag fish population and the optimal individual of the fitness value of the sardine population in the initial state, and respectively recording the optimal individual as the flag elegans and the optimal individual as injured sardines;

s46, updating the self position of the flag fish population according to the position of the Elfin flag fish and the position of the injured sardine:

wherein the content of the first and second substances,andrespectively representing the positions of the current Elaphe arguin and the injured sardine, wherein r is a random number uniformly distributed between 0 and 1, and lambda is_iIs a dynamically changing coefficient, whose expression is:

wherein N is_SFAnd N_SRespectively representing the number of the flag fishes and the sardines;

s47, the sardine population updates the position according to the attacking force of the flag fish and the orientation of the Elfin flag fish:

wherein the content of the first and second substances,the current position of the sardine is shown, the AP shows the attack force of the flag fish, and the attack force is linearly reduced after each attack, and the calculation formula is as follows:

AP＝A*(1-2Itr*ε)

wherein A and epsilon are attack force change coefficients of the flag fish, and Itr is the current iteration number;

s48, recalculating the fitness values of the flag fishes and the sardines and comparing the fitness values, wherein if the fitness values of the sardines are better than the flag fishes, the flag fishes occupy the positions of the sardines and remove the sardines from the population, and the method can be specifically expressed as follows:

if f(S_i)<f(SF_i)

wherein the content of the first and second substances,is the position of the flag fish at present,as the current sardine position, S_iIs the fitness of sardines to, SF_iThe fitness value of the flag fish is obtained;

s49, judging whether the maximum iteration number is reached, if so, outputting the position of the current sailfish elite, namely the optimal parameter of the TSVR, otherwise, returning to the step S46;

s5, verifying the trained TSVR model based on the test set, if the verification is passed, executing the step S6, otherwise, returning to the step S3, specifically:

s51, selecting corresponding data from the test set to input the trained TSVR model according to the set input data sequence to obtain a corresponding verification predicted value;

s52, carrying out reverse normalization processing on the verification prediction value to obtain a verification prediction result;

s53, evaluating the verification prediction result based on the corresponding real load data in the test set, if the evaluation value is less than or equal to the preset threshold value, the verification is passed, and step S6 is executed, otherwise, the step S4 is returned, wherein the evaluation is carried out by adopting a mode of calculating the average absolute percentage error, and the method comprises the following steps:

wherein MAPE is the mean absolute percentage error, K is the number of sample points to be predicted,and y_iRespectively verifying a predicted value and a true value of the load data, if the MAPE value obtained by calculation is less than or equal to a preset threshold value, the verification is passed, otherwise, the verification is not passed;

and S6, inputting the relevant data of the actual day to be predicted into the trained TSVR model according to the set input data sequence to obtain the predicted value of the power load of the actual day to be predicted.

The method is applied in this embodiment, a specific load prediction process of the method is shown in fig. 2, an implementation platform of the embodiment is intel (r) core (tm) i5-8300H CPU @2.30GHz, and RAM 8GB, and an experimental environment is based on a Matlab platform. In this embodiment, the method of the present invention is verified by using the real data of 2009, 2 month, 20 days to 5 month, 10 days in a certain city in east china, and the data set characteristics include 24-point historical load data, meteorological data and date data types, wherein the meteorological data includes a highest daily temperature, an average daily temperature, a lowest daily temperature and a relative humidity, and the date data types include a working day, a weekend and a holiday.

The specific processes of step S1 and step S2 include:

the historical load data and the meteorological data in the sample are normalized, and corresponding date type data are quantized to be in a [0,1] range, and the specific method is as shown in table 1:

TABLE 1

In Table 1, L_minAnd L_maxMinimum and maximum values of the load data, respectively; w' is the normalized meteorological data; w_minAnd W_maxRespectively, the minimum value and the maximum value of the meteorological data of the same category.

And then, dividing the original sample data into a training set and a test set, wherein the training set data selects sample data from 20 days to 30 days in 2 months in 2009, the rest sample data is used as the test set, and a 24-point load prediction experiment is carried out on 10 days in 5 months in 2009.

The specific process of step S3 includes:

firstly, constructing an input data sequence:

selecting the date type and meteorological data and load data of two days before the day to be predicted, the date type and meteorological data and load data of one day before the day to be predicted, the date type and meteorological data of the day to be predicted as an input sequence of a TSVR model, and setting the ith input sequence X as the input sequence of the TSVR model_iCan be expressed as

X_i＝[D_i-2 W_i-2 L_i-2 D_i-1 W_i-1 L_i-1 D_i W_i]

In the formula, D_i-2、W_i-2、L_i-2Respectively representing the date type, the weather type and the load data of two days before the day to be predicted; d_i-1、W_i-1、L_i-1Respectively representing the date type, the weather type and the load data of the day before the day to be predicted; d_i-1、W_i-1Respectively representing the date type and weather data of the day to be predicted.

Secondly, establishing a TSVR regression model according to input data, comprising the following steps:

1. establishing a regression function expression of the TSVR:

the regression function of the TSVR is the average of its upper and lower bound decision functions and can be expressed as

f(x)＝(f₁(x)+f₁(x))/2

In the formula (f)₁(x) And f₂(x) Are respectively provided withIs the upper and lower bound decision function of TSVR, and its specific expression is as follows

In the formula, w₁And w₂Are respectively a decision function f₁(x) And f₂(x) The weight coefficient of (a); b₁And b₂Is the bias term.

2. Solving the regression function expression of the TSVR:

converting the solving problem of the weight coefficients and the bias terms of the upper and lower bound decision functions of the TSVR into the following two small convex quadratic programming problems:

s.t.Y-(Xw₁+eb₁)≥eε₁-ξ₁,ξ₁≥0

s.t.(Xw₂+eb₂)-Y≥eε₂-ξ₂,ξ₂≥0

in the formula, C₁And C₂Is a penalty factor; input matrix X ═ X₁,x₂,…,x_n]^T(ii) a Output matrix Y ═ Y₁,y₂,…,y_n]^T；ε₁,ε₂Is an insensitive loss factor; xi₁And xi₂Is a relaxation variable; e is a column vector with elements all 1.

Lagrange operators alpha and beta are introduced, and weight coefficients and bias terms of upper and lower bound decision functions of the TSVR are solved according to the KKT condition

Wherein G ═ X e]，f＝Y-eε₁，h＝Y+eε₂. Thus, a regression function of the TSVR can be obtained

The specific process of step S4 is:

firstly, determining an optimization target and a fitness function of an algorithm:

determining a penalty parameter C for a TSVR to be optimized₁、C₂Insensitive loss factor ε₁And ε₂And a gaussian kernel parameter sigma, the position of each flag fish corresponding to a set of parameters of the TSVR;

and determining the fitness function of the SFO, wherein the method comprises the following steps:

wherein M is the number of training set samples;and y_kThe actual output value and the expected output value of the training sample are respectively.

Secondly, optimizing TSVR parameters by the SFO algorithm, wherein the specific process comprises the following steps:

1. initializing the SFO algorithm:

setting a feasible field of a parameter to be optimized, wherein a penalty parameter C₁、C₂∈[0.1,100]Insensitive loss factor ε₁And ε₂∈[0.001,1]The Gaussian kernel parameter σ ∈ [0.01, 1]0]。

Randomly initializing the positions of the flag fish population and the sardine population according to the feasible region of the parameter to be optimized, setting the maximum iteration number M of the SFO algorithm to be 50, and respectively setting the number of the flag fish and the number of the sardine to be N_S15 and N_SF35, the attack coefficient A of the flag fish is 4, and the epsilon is 0.001;

calculating initial fitness values of the flag fish population and the sardine population; and screening the optimal individual of the fitness value of the flag fish population and the optimal individual of the fitness value of the sardine population in the initial state, and respectively recording the individuals as the elite flag fish and the injured sardine.

2. SFO algorithm iteration optimizing process:

the sailfish population updates the position of the sailfish population according to the position of the sailfish Elite and the position of the injured sardine, and the specific calculation formula is as follows

Wherein the content of the first and second substances,andrespectively showing the positions of the current Elite flagfish and the injured sardine; r is a random number uniformly distributed between 0 and 1; lambda [ alpha ]_iIs a dynamically changing coefficient expressed as

In the formula, N_SFAnd N_SRepresenting the number of flag fish and sardine, respectively.

The sardine population can update the position according to the attacking force of the flag fish and the direction of the Elaeagnus guichenensis, and the position updating formula is

In the formula (I), the compound is shown in the specification,indicating the current sardine location; AP represents the attack force of the flag fish, and the attack force linearly decreases after each attack, and the calculation formula is

AP＝A*(1-2Itr*ε)

Wherein A and epsilon are the attack force variation coefficient of the flag fish; itr is the current iteration number.

Recalculating the fitness values of the flag fishes and the sardines and comparing the fitness values, wherein if the fitness values of the sardines are better than the flag fishes, the flag fishes occupy the positions of the sardines and remove the sardines from the population, which can be specifically expressed as

if f(S_i)<f(SF_i)

Judging whether the maximum iteration times is reached, if so, outputting the position of the current Eleutherococcus elegans and ending the SFO algorithm, namely the optimal parameter of the TSVR, otherwise, repeating the iteration optimization process;

in step S5, a TSVR model is used to perform a load prediction experiment on 24-point load data in the test set at 5, 10 months and 2009 to obtain a predicted value, and the predicted value is subjected to inverse normalization to obtain a final prediction result, which is shown in fig. 3.

And finally, evaluating the short-term power load prediction result of the method, and adopting the average absolute percentage error MAPE as an evaluation standard:

k is the number of sample points to be predicted;and y_iRespectively, the predicted value and the true value of the load data.

In order to verify the prediction effect of the SFO-TSVR, a comparison experiment is carried out with a PSO-TSVR model and a WOA-TSVR model, the prediction result is shown in figure 4, and it can be seen that the prediction value of the SFO-TSVR model is more consistent with an actual load curve on the whole, so that the SFO algorithm has better optimization performance than the PSO algorithm and the WOA algorithm, and the prediction performance of the model is effectively improved.

In order to further verify the prediction result of the SFO-TSVR, LSTM and SVR models are selected as comparison models, the comparison results and the evaluation indexes of the models are respectively shown in figure 5 and table 2, and the combination of figure 5 and table 2 shows that the method has the minimum prediction error, compared with the TSVR model and the LSTM model, the average absolute percentage error is respectively reduced by 52.85% and 41.30%, and the effectiveness of the method is proved.

TABLE 2

Model (model)	LSTM	SVR	TSVR	SFO-TSVR
					MAPE/％	3.24	4.40	4.03	1.90

In conclusion, the SFO-TSVR short-term power load prediction model is created, the TSVR model is selected by using the SFO algorithm with high convergence rate and high optimization precision, and compared with other model parameter determination modes, the efficiency is higher and more stable;

the short-term power load prediction model of the SFO-TSVR, which is established by the invention, is established on the basis of empirical risk minimization, so that the method can be suitable for solving large-scale data, overcomes the defect that support vector regression is not good at processing a large amount of data, and has better performance on training small sample data;

the SFO-TSVR model created by the method has higher prediction precision relative to models such as support vector regression, long-term and short-term memory networks and twin support vector regression.