An ISSA-HKELM-based short-term load forecasting method

Abstract

The present application relates to a kind of short-term load prediction method based on ISSA-HKELM, the method includes: first, in view of the defect of kernel extreme learning machine KELM, combined with Gaussian kernel function and polynomial kernel function, construct the hybrid kernel extreme learning machine HKELM with stronger generalization ability;Second, in view of the problem that sparrow search algorithm is easy to fall into local extremum, adaptive t distribution strategy and dynamic adaptive weight are introduced to improve sparrow search algorithm;Third, the improved sparrow search algorithm ISSA is used to optimize the parameter of hybrid kernel extreme learning machine HKELM and establish ISSA-HKELM prediction model;Finally, short-term load prediction is carried out using the established ISSA-HKELM model.Compared with prior art, the present application has good prediction accuracy and robustness and the like advantages.

Description

Technical Field

[0001] This invention relates to a short-term load forecasting method, and more particularly to a short-term load forecasting method based on ISSA (Improved Sparrow Search Algorithm)-HKELM (Optimized Hybrid Kernel Extreme Learning Machine). Background Technology

[0002] Load forecasting is a crucial foundation for energy planning, economic operation, and energy management. It typically includes long-term, medium-term, and short-term load forecasting. Short-term load forecasting generally refers to predicting load for the next day or week. A key characteristic of short-term load forecasting is its significant susceptibility to factors such as weather, equipment condition, and major social events, making accurate short-term load forecasting quite challenging. Industrial, residential, and utility electricity loads exhibit vastly different characteristics, and electricity loads often experience significant fluctuations and seasonality due to weather changes. Accurate electricity load forecasting is fundamental to the development of power system expansion, operation, and maintenance plans.

[0003] In the field of short-term power load forecasting, forecasting methods are mainly divided into traditional statistical methods and machine learning methods. Traditional methods mainly include time series analysis, regression analysis, and Kal's filtering. These methods mostly rely on fitting historical data to predict future load trends. While simple, they cannot reflect the nonlinear characteristics of the load. Machine learning methods mainly include backpropagation neural networks, support vector machines (SVM), and extreme learning machines (ELM). Although these methods have higher prediction accuracy than traditional methods, they also suffer from complex processes, poor stability, and difficulty in parameter adjustment. Kernel Extreme Learning Machine (KELM) is a novel single-hidden-layer feedforward neural network algorithm based on kernel functions, which has shown good performance in handling nonlinear regression. Traditional KELM algorithms mostly use a single kernel function, resulting in weak generalization ability and predictive performance. Sparrow Search, as a novel swarm intelligence optimization algorithm, has advantages such as simple parameter setting and fast convergence speed. However, Sparrow Search suffers from reduced population diversity and susceptibility to local optima in the later stages of iteration. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a short-term load forecasting method based on ISSA-HKELM with good forecasting accuracy and robustness.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] According to one aspect of the present invention, a short-term load forecasting method based on ISSA-HKELM is provided, the method comprising:

[0007] First, to address the shortcomings of the Kernel Extreme Learning Machine (KELM), a Hybrid Kernel Extreme Learning Machine (HKELM) with stronger generalization ability is constructed by combining Gaussian kernel function and polynomial kernel function.

[0008] Secondly, to address the problem that the sparrow search algorithm is prone to getting trapped in local extrema, an adaptive t-distribution strategy and dynamic adaptive weights are introduced to improve the sparrow search algorithm;

[0009] The improved Sparrow Search Algorithm (ISSA) was used again to optimize the parameters of the Hybrid Kernel Extreme Learning Machine (HKELM) and an ISSA-HKELM prediction model was established.

[0010] Finally, the established ISSA-HKELM model was used for short-term load forecasting.

[0011] As a preferred technical solution, the input and output variables of the ISSA-HKELM prediction model are selected as follows:

[0012] The model inputs include historical load data for the day before the forecast, the previous day's highest, lowest, average temperature, relative humidity, and rainfall, as well as the current day's highest, lowest, average temperature, relative humidity, and rainfall.

[0013] The load data for the forecast date will be used as the output.

[0014] As a preferred technical solution, the Hybrid Kernel Extreme Learning Machine (HKELM) effectively solves the problem of weak generalization ability of kernel extreme learning machines.

[0015] As a preferred technical solution, the adaptive t-distribution strategy is specifically as follows:

[0016] The sparrow's position is updated using an adaptive t-distribution as shown in the following formula.

[0017]

[0018] Where x i Let i be the position of the i-th sparrow. t represents the position of the sparrow after mutation t; t(k) represents the t-distribution with parameter degrees of freedom k for iteration number.

[0019] As a preferred technical solution, when the number of iterations k in the early stage is small, the t-distribution is similar to the Cauchy distribution mutation. At this time, the t-distribution operator is likely to obtain a large value, and the step size of the position mutation is large, giving the algorithm a good global search capability. In the middle of the iteration, the t-distribution changes from the Cauchy distribution mutation to the Gaussian distribution mutation. At this time, the t-distribution operator is likely to obtain a relatively moderate value, allowing the algorithm to take into account both global and local search capabilities. When the number of iterations k in the later stage is large, the t-distribution is similar to the Gaussian distribution mutation. At this time, the t-distribution operator is likely to obtain a small value, and the step size of the position mutation is small.

[0020] As a preferred technical solution, the dynamic adaptive weights are specifically as follows:

[0021] The concept of inertial weight is introduced into the position update formula of the discoverer in the game of Sparrow and improved by introducing a dynamic weight factor ω into the position update method of the discoverer.

[0022] As a preferred technical solution, the value of the dynamic weighting factor ω is related to the maximum number of iterations K and the current number of iterations k, and decreases as the current number of iterations k increases; in the early stage of iteration, when the number of iterations is small, the value of ω is large; in the later stage of iteration, when the number of iterations increases, the value of ω is small.

[0023] As a preferred technical solution, the calculation formula for the dynamic weighting factor ω and the improved discoverer position update method are as follows:

[0024]

[0025]

[0026] in Let be the globally optimal solution in the j-th dimension of the previous generation, and k represent the current iteration number; j = (1, 2, ..., d); Let represent the position of the i-th sparrow in the j-th dimension during the k-th iteration; K represents the maximum number of iterations; z is a random number in (0,1]; R2 represents the warning value and R2∈[0,1]; ST represents the safety threshold and ST∈[0.5,1]; Q represents a random value that follows a normal distribution; L is a 1×d matrix with all internal elements being 1.

[0027] As a preferred technical solution, the specific process of optimizing the parameters of the hybrid kernel extreme learning machine HKELM using the improved sparrow search algorithm ISSA is as follows:

[0028] Step 1: Divide the distribution network data into training and testing sets and perform normalization processing;

[0029] Step 2: Initialize parameters: sparrow population size N, maximum number of iterations K, safety threshold ST, proportion of discoverers PD, proportion of scouts SD, t-distribution mutation probability p, and the optimization interval of model parameters;

[0030] Step 3: Calculate the fitness value, take the predicted value yi' obtained by the model learning from the training samples and the true value y i the mean square error of is taken as the fitness value;

[0031] Step 4: Calculate the fitness values, sort them, obtain the optimal and worst individuals and their corresponding positions, select part of the sparrows with better fitness values as discoverers, and the rest as followers;

[0032] Step 5: Update the positions of followers, sentinels and discoverers respectively according to their respective position update formulas, calculate the corresponding fitness value after update, compare it with the current best fitness value, and update the global optimal information;

[0033] Step 6: If rand<p, perform t-distribution mutation operation on sparrows, where rand represents a random number between 0 and 1, and the value of p is 0.5; otherwise, return to Step 3;

[0034] Step 7: Calculate the fitness value of the sparrow after t-distribution mutation, if the fitness value of the new solution after t-distribution mutation is better than the global optimal value, replace the original global optimal value with the new value after mutation, otherwise keep the original one;

[0035] Step 8: Termination condition: if the maximum number of iterations is reached, output the model parameters to obtain the optimal ISSA-HKELM model; otherwise, return to Step 3.

[0036] As a preferred technical solution, root mean square error, average absolute percentage error, average absolute error and goodness of fit are selected as the evaluation indicators of the ISSA-HKELM model.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1) The present invention uses the optimization capability of ISSA (Improved Sparrow Search Algorithm) to optimize the parameters of HKELM, and uses HKELM to mine useful information in load data, so as to better fit the load data and complete ultra-short-term power load prediction accurately and efficiently.

[0039] 2) To address the weak generalization ability of the KELM model, this invention constructs a Hybrid Kernel Extreme Learning Machine (HKELM) with stronger generalization ability by combining Gaussian kernel function and polynomial kernel function. To address the issue that the prediction performance of the HKELM model is determined by the model parameters, a sparrow search algorithm is used to optimize the selection of its parameters. Furthermore, to address the problem that the sparrow search algorithm suffers from reduced population diversity and susceptibility to local optima in the later stages of iteration, a short-term power load prediction model based on an improved sparrow search algorithm to optimize the HKELM is proposed, enhancing the stability of the model prediction and improving the prediction accuracy. Attached Figure Description

[0040] Figure 1 Topology diagram of the kernel-based extreme learning machine;

[0041] Figure 2 This is a graph of the polynomial kernel function.

[0042] Figure 3 The graph shows the Gaussian kernel function.

[0043] Figure 4 A graph of mixed kernel functions;

[0044] Figure 5 Plots of the t-distribution, Gaussian distribution, and Cauchy distribution functions;

[0045] Figure 6 Flowchart of the ISSA-HKELM model;

[0046] Figure 7 This is a comparison chart of the prediction results from different models. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] This invention discloses a short-term power load forecasting method based on an improved sparrow search algorithm and an optimized hybrid kernel extreme learning machine (HKELM). First, a hybrid kernel extreme learning machine (HKELM) with stronger generalization ability is constructed by combining Gaussian and polynomial kernel functions. Then, to address the problem of the sparrow search algorithm easily getting trapped in local optima, an adaptive t-distribution strategy and dynamic adaptive weights are introduced to improve the sparrow search algorithm. Finally, an ISSA-HKELM model is established for short-term load forecasting, and the prediction results are compared and analyzed with those of LSTM, DBN, KELM, HKELM, PSO-HKELM, and SSA-HKELM models. The method includes the following steps:

[0049] A. Selection of input and output variables for the prediction model:

[0050] Electricity load data itself has dynamic characteristics. Historical load data is divided into 96 time points (one point every 15 minutes) on a daily basis to represent the load's own fluctuation pattern. The historical load data of the day before the forecast date, the highest temperature, lowest temperature, average temperature, relative humidity and rainfall of the previous day, and the highest temperature, lowest temperature, average temperature, relative humidity and rainfall of the current day are used as the inputs to the model (106 inputs). The load data of the forecast day is used as the output for model training and prediction (96 outputs).

[0051] B. Hybrid Kernel Extreme Learning Machine (HKELM)

[0052] Kernel Extreme Learning Machine (KELM), a novel single-hidden-layer feedforward neural network algorithm based on kernel functions, demonstrates excellent performance in handling nonlinear regression. The output of this network can be expressed as:

[0053] f(x)=h(x)β=Hβ (1)

[0054] In the formula, x represents the input data, f(x) represents the predicted output of the KELM model, h(x) represents the mapping function of the hidden layer, and H represents the matrix generated by the kernel function mapping of the input samples, which consists of elements x i The kernel matrix of KELM is composed of β, which represents the weights between the hidden layer and the output layer.

[0055]

[0056] KELM uses the kernel matrix Ω ELM Replaces HH in ELM T And the input samples are mapped to a high-dimensional hidden layer feature space through a kernel function, where K(x) in the formula i ,x j () represents a kernel function. Common kernel functions are mainly divided into the following four types:

[0057] 1) Gaussian kernel function

[0058] K(x i ,x j )=exp(-||x i -x j || 2 / (2σ 2 (3)

[0059] 2) Polynomial kernel function

[0060] K(x i ,x j )=(m(x i ·x j )+n) d ;d=1,2,…,N (4)

[0061] 3) Linear kernel function

[0062] K(x i ,x j )=x i ·x j (5)

[0063] 4) Perceptron kernel function

[0064] K(x i ,x j )=tanh(β(x i ·x j (6)+b)

[0065] KELM topology diagram Figure 1 As shown, x is the input vector of the model, x1, x2, ... x N To prevent overfitting of the model, the kernel matrix Ω is used as the training set samples. ELM Add an identity matrix I0 and a penalty coefficient C to the main diagonal, and then calculate β. This makes the kernel extreme learning machine have better generalization ability. The output and output weight vector are shown in equations (7) and (8).

[0066]

[0067] β=(I0 / C+Ω ELM ) -1 T (8)

[0068] In the formula, T is the target output matrix.

[0069] The performance of KELM is determined by the kernel function. This paper combines the polynomial kernel function and the Gaussian kernel function as the kernel functions of KELM to improve the model's performance. The polynomial kernel function is a global kernel function with poor local learning ability. The polynomial kernel function curve shows that, for example, in a test case X = 0.5, m = 1, n = 1, and d takes values ​​of 1, 2, 3, and 4... Figure 2 It can be seen that the global characteristics of the polynomial kernel function are reflected in the greater influence on points that are far away from the test point.

[0070] The Gaussian kernel function is a local kernel function with excellent local learning capabilities, but it cannot accurately and effectively predict samples beyond a certain range. (Based on the Gaussian kernel function curve...) Figure 3 It can be seen that the test point X = 0.5, and σ takes values ​​of 0.1, 0.2, 0.3 and 0.4. This shows that the local characteristics of the Gaussian kernel function are reflected in the fact that the points closer to the test point have a greater influence.

[0071] Hybrid kernel functions can combine the advantages of both, therefore, a hybrid kernel function is formed by linearly combining two kernel functions:

[0072]

[0073] Based on the hybrid kernel function curve Figure 4 It can be seen that the test point X = 0.5, λ takes values ​​of 0.1, 0.2, 0.3 and 0.4, σ takes 0.1 in the Gaussian kernel function, and m = 1, n = 1 and d = 2 in the polynomial kernel function. The mixed kernel function not only affects the sample points around the test point, but also has a certain influence on the sample points at a certain distance from the test point. Therefore, the mixed kernel function effectively combines the advantages of the Gaussian kernel function and the polynomial kernel function, and makes up for the shortcomings of the single kernel function.

[0074] Where σ, m, n, and d are the hybrid kernel parameters, and λ is the weight coefficient of the polynomial kernel function, where d = 2. An improved sparrow search algorithm is used to optimize σ, m, n, λ, and the penalty coefficient C when training the Hybrid Kernel Extreme Learning Machine (HKELM).

[0075] C. Improve the sparrow search algorithm to optimize the hybrid kernel extreme learning machine model (ISSA-HKELM)

[0076] The Sparrow Search Algorithm (SSA) is a novel intelligent optimization algorithm proposed by Xue et al. in 2020. This algorithm is primarily inspired by the predation and anti-predation behaviors of sparrows. The sparrow set matrix is ​​as follows:

[0077] X = [x1, x2, ..., x N ] T,x i =[x i,1 ,x i,2 ,…,x i,d (10)

[0078] In the formula, N represents the number of sparrows, i = (1,2,…,N), and d represents the dimension of the variable.

[0079] The fitness matrix of a sparrow is represented as follows:

[0080] F x =[f(x1),f(x2),…,f(x N )] T (11)

[0081] f(x i )=[f(x i,1 ),f(x i,2 ),…,f(x i,d (12)

[0082] In the formula, F x Each value in the table represents the fitness value of an individual. The sparrow population is divided into discoverers, followers, and vigilants. In each iteration, a subset of sparrows with relatively high fitness values ​​are selected as discoverers, generally accounting for 10% to 20% of the population. They are mainly responsible for leading the population to places with food. The rest are followers, while vigilants are randomly selected from the entire population, representing 10% to 20%.

[0083] The location of the discoverer is updated as follows:

[0084]

[0085] In the formula, k represents the current iteration number; j = (1, 2, ..., d); Let Ri represent the position of the i-th sparrow in the j-th dimension during the k-th iteration; K represents the maximum number of iterations; z is a random number in (0,1]; R2 represents the warning value and R2 ∈ [0,1]; ST represents the safety threshold and ST ∈ [0.5,1]; Q represents a random value following a normal distribution; L is a 1×d matrix with all internal elements being 1. When R2 < ST, it indicates that the warning value is lower than the safety value, the sparrow population is in a relatively safe place with no natural predators nearby, and the discoverer can conduct a wide search of the area; when R2 ≥ ST, it indicates that the warning value exceeds the safety value, there is danger around the sparrow population, and the entire population needs to migrate to a safe place.

[0086] The formula for updating the position of followers is as follows:

[0087]

[0088] In the formula, Xworst Indicates the current worst position globally; X P A represents the optimal location of the discoverer; A represents a 1×d matrix whose internal elements are all 1 or -1; where A + =A T (AA T ) -1 When i > N / 2, it indicates that the i-th follower has poor fitness and needs to go to other areas to find food.

[0089] The position update formula for the vigilant is:

[0090]

[0091] In the formula, X best β represents the current global optimal position; β represents the direction of the sparrow's movement, which is a random number following a normal distribution; m is a random value in [-1, 1]; f i f represents the fitness value of the current individual sparrow; g and f w These represent the current global optimal and worst fitness values, respectively. To prevent (f i -f w If f + ε = 0 makes the denominator 0, then ε is defined as the smallest constant, which is set to 10E-8 in this paper. i >f g When f indicates that the sparrow's current fitness value exceeds the current optimal fitness value, the sparrow is located on the edge of the population and is vulnerable to danger; when f i =f g When the value is 0, it indicates that the sparrow's current fitness value is the optimal fitness value at present. The sparrow in the center of the population has sensed the danger and needs to move to the location of other sparrows.

[0092] To address the shortcomings of the Sparrow Search Algorithm (SSA), an Improved Sparrow Search Algorithm (ISSA) is proposed. Based on SSA, an adaptive t-distribution strategy and dynamic adaptive weights are introduced to avoid the algorithm getting trapped in local optima and improve the convergence speed and accuracy of the algorithm.

[0093] (1) Adaptive t-distribution strategy

[0094] The t-distribution is short for the student distribution. The shape of its distribution function curve is determined by the value of the parameter degrees of freedom, n, and can be obtained from the function distribution graph. Figure 5It can be seen that when the degree of freedom n is smaller, the t-distribution curve is flatter in the middle, with higher double tails, and is smoother overall. When the degree of freedom n = 1, the t-distribution is a Cauchy distribution, i.e., t(n = 1) → C(0, 1). When the degree of freedom n is larger, the t-distribution curve is steeper in the middle, and is more precipitous overall. When the degree of freedom n is infinitely large, the t-distribution is a Gaussian distribution, i.e., t(n = ∞) → N(0, 1).

[0095] The sparrow's position is updated using an adaptive t-distribution as shown in the following formula:

[0096]

[0097] In the formula, x i Let i be the position of the i-th sparrow. Let t(k) represent the position of the sparrow after mutation t; t(k) represents the t-distribution with parameter degrees of freedom k iterations. Information about the current sparrow population is fully utilized in this formula. When the initial iteration number k is small, the t-distribution resembles a Cauchy distribution mutation, where the t-distribution operator is likely to obtain a large value, and the step size for position mutation is large, giving the algorithm good global search capabilities. In the middle of the iteration, the t-distribution transforms from a Cauchy distribution mutation to a Gaussian distribution mutation, where the t-distribution operator is likely to obtain a relatively moderate value, allowing the algorithm to consider both global and local search capabilities. When the later iteration number k is large, the t-distribution resembles a Gaussian distribution mutation, where the t-distribution operator is likely to obtain a small value, and the step size for position mutation is small, giving the algorithm good local search capabilities. This strategy is beneficial for improving the algorithm's overall search capabilities.

[0098] (2) Dynamic adaptive weights

[0099] As can be seen from the discovery's position update formula, at the very beginning of the algorithm's iterations, the discovery approaches the global optimum, resulting in a small search space and a tendency to get trapped in local optima. Therefore, this chapter considers introducing the concept of inertial weights into the discovery's position update formula in the sparrow problem and improving it by introducing a dynamic weight factor ω. The value of ω is related to the maximum number of iterations K and the current number of iterations k, and decreases as the current number of iterations k increases. In the early stages of iteration, when the number of iterations is small, the value of ω is large, which allows for effective global search; in the later stages of iteration, when the number of iterations increases, the value of ω is small, which allows for effective local search.

[0100] Simultaneously, the global optimal solution from the previous generation is incorporated into the discoverer's position update formula, allowing both the previous discoverer's position and the previous generation's global optimal solution to simultaneously influence the current discoverer's position. This prevents the algorithm from getting trapped in local optima. The calculation formula for the weight coefficient ω and the improved discoverer position update method are shown below:

[0101]

[0102]

[0103] In the formula, represents the global optimal solution of the j-th dimension in the previous generation; rand represents a random number between 0 and 1.

[0104] As Figure 6 shown, the prediction model of hybrid kernel extreme learning machine optimized by improved sparrow search algorithm (ISSA-HKELM) adopts the improved sparrow search algorithm (ISSA) to optimize the penalty coefficient C, hybrid kernel parameters σ, m and n, and weight coefficient λ of the hybrid kernel extreme learning machine (HKELM), so as to obtain the optimal model. The specific steps are as follows:

[0105] 1) Divide the load data into a training set and a test set and perform normalization processing.

[0106] 2) Initialize parameters, including sparrow population size N, maximum number of iterations K, safety threshold ST, proportion of discoverers PD, proportion of spotters SD, t-distribution mutation probability p, and the optimization interval of model parameters, etc.

[0107] 3) Calculate the fitness value, the predicted value y' obtained by the model learning from training samples i and the real value y i the mean square error (MSE) of is used as the fitness value, and the MSE calculation formula is as follows.

[0108]

[0109] 4) Calculate and sort the fitness values to generate the optimal and worst individuals and their corresponding positions. Select some sparrows with better fitness values as discoverers, and the remaining as followers.

[0110] 5) Update the positions of followers, alert sparrows and discoverers according to formulas (13), (14) and (15) respectively, calculate the corresponding fitness value after updating, compare it with the current best fitness value, and update the global optimal information.

[0111] 6) If rand<p, perform t-distribution mutation operation on sparrows according to formula (16). Otherwise, go to step 3.

[0112] 7) Calculate the fitness value of sparrows after t-distribution mutation. If the fitness value of the new solution after t-distribution mutation is better than the global optimal value, replace the previous global optimal value with the new value after mutation; otherwise, keep the previous one.

[0113] 8) Termination condition. If the maximum number of iterations has been reached, output the model parameters to obtain the optimal HKELM model; otherwise, return to step 3.

[0114] 9) Use the established hybrid kernel extreme learning machine model to predict short-term power load.

[0115] Root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), and goodness of fit (R²) were selected. 2 () is used as an evaluation metric for the model.

[0116]

[0117]

[0118]

[0119]

[0120] This invention provides an example analysis of the short-term load forecasting model. Based on the historical load of a certain region from January 1, 2012 to January 17, 2015, at 96 time points daily, and incorporating relevant factors influencing load, including maximum temperature, minimum temperature, average temperature, relative humidity, and rainfall, the multidimensional data is divided into input and output nodes for load forecasting. The data is divided into training data and test data. The training data is from January 1, 2012 to January 16, 2015, and the test data is the data from the last day. The prediction results of the seven models are shown in the figure below. Figure 7 As shown in Table 1, the evaluation index values ​​of the prediction results are as follows.

[0121] Table 1

[0122] RMSE 582.93 2126.80 212.87 234.69 103.20 78.68 61.71 MAE 487.44 1695.07 186.69 212.90 82.62 64.62 49.82 MAPE 7.27％ 26.32％ 2.53％ 2.9758％ 1.22％ 0.94％ 0.77％ <![CDATA[R 2 ]]> 0.95502 0.83482 0.99643 0.99632 0.99649 0.99746 0.99791

[0123] As can be seen from the evaluation index values ​​in Table 1, the root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) of the ISSA-HKELM model proposed in this embodiment are all reduced to some extent compared with the LSTM model, DBN model, KELM model, HKELM model, PSO-HKELM model, and SSA-HKELM model. The smaller the error index, the better the prediction effect; and the goodness of fit (R²) is also significantly reduced. 2 The highest value is indicated by the fit, where a higher fit indicates a more accurate prediction. In summary, the method in this embodiment effectively removes the randomness of prediction, reduces the error of the predicted values, and greatly improves the accuracy of prediction, further demonstrating the advantages of the ISSA-HKELM model in the field of short-term power load forecasting.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A short-term load forecasting method based on ISSA-HKELM, characterized in that, Applied to the field of power grid dispatching, the parameters of HKELM are optimized by utilizing the optimization capabilities of ISSA, and useful information within the load data is extracted through HKELM to achieve a better fit to the load data. The prediction method includes: First, to address the shortcomings of the Kernel Extreme Learning Machine (KELM), a Hybrid Kernel Extreme Learning Machine (HKELM) with stronger generalization ability is constructed by combining Gaussian kernel function and polynomial kernel function. Secondly, to address the problem of the sparrow search algorithm easily getting trapped in local optima, an adaptive approach is introduced. t The sparrow search algorithm is improved by using a distribution strategy and dynamic adaptive weights; The improved Sparrow Search Algorithm (ISSA) was used again to optimize the parameters of the Hybrid Kernel Extreme Learning Machine (HKELM) and an ISSA-HKELM prediction model was established. Finally, the established ISSA-HKELM model was used for short-term load forecasting. The input and output variables of the ISSA-HKELM prediction model are selected as follows: The model inputs include historical load data for the day before the forecast, the previous day's highest, lowest, average temperature, relative humidity, and rainfall, as well as the current day's highest, lowest, average temperature, relative humidity, and rainfall. Use the load data for the forecast day as the output; The improved Sparrow Search Algorithm (ISSA) optimizes the parameters of the Hybrid Kernel Extreme Learning Machine (HKELM) as follows: Step 1: Divide the distribution network data into training and testing sets and perform normalization processing; Step 2: Initialize parameters, sparrow population size Maximum number of iterations K Safety threshold Discoverer ratio The proportion of investigators , t Distribution of Probability of Variation p And the optimization range of model parameters; Step 3: Calculate the fitness value, which is the predicted value obtained by the model learning from the training samples. and the true value The mean square error is used as the fitness value; Step 4: Calculate and sort the fitness values ​​to generate the best and worst individuals and their corresponding positions. Select some sparrows with better fitness values ​​as discoverers and the rest as followers. Step 5: Update the positions of the followers, vigilants, and discoverers according to their respective position update formulas, calculate the corresponding fitness values ​​after the update and compare them with the current best fitness values, and update the global optimal information. Step 6: If sparrows t Distribution mutation operation, where Represents a random number between 0 and 1. p The value is 0.5; otherwise, return to step 3. Step 7: Calculate the number of sparrows. t The fitness value after distribution variation, if after t If the fitness value of the new solution after the mutation operation is better than the global optimum, then the new value after mutation is used to replace the previous global optimum; otherwise, it is retained. Step 8: Termination condition: If the maximum number of iterations has been reached, output the model parameters to obtain the optimal ISSA-HKELM model; otherwise, return to step 3. The adaptive t The specific distribution strategy is as follows: Adaptive method for sparrow position t The distribution is updated as shown in the following formula. in x i For the first i The location of a single sparrow; Indicates the process t The location of the mutated sparrow; t ( k () indicates that the parameter degrees of freedom are the number of iterations. k of t distributed; When the number of previous iterations k When smaller, t The distribution is similar to the variation of the Cauchy distribution, at this time t The distribution operator has a high probability of obtaining a large value, and the step size for position mutation is relatively large, giving the algorithm good global search capabilities; in the middle of the iteration, t The distribution changes from a Cauchy distribution to a Gaussian distribution at this point. t The distribution operator likely takes a relatively moderate value, allowing the algorithm to consider both global and local search capabilities; when the number of later iterations... k When it is large, t The distribution resembles the variation of a Gaussian distribution, at this time... t The distribution operator is likely to obtain a small value, and the step size for position variation is small; The model inputs include historical load data at 96 time points the day before the forecast, the previous day's highest, lowest, average temperature, relative humidity, and rainfall, as well as the current day's highest, lowest, average temperature, relative humidity, and rainfall.

2. The short-term load forecasting method based on ISSA-HKELM according to claim 1, characterized in that, The proposed Hybrid Kernel Extreme Learning Machine (HKELM) effectively addresses the issue of weak generalization ability in kernel extreme learning machines.

3. The short-term load forecasting method based on ISSA-HKELM according to claim 1, characterized in that, The dynamic adaptive weights are specifically: The concept of inertia weight is introduced into the position update formula for the discoverer in a sparrow game and improved by introducing a dynamic weight factor into the discoverer's position update method. .

4. The short-term load forecasting method based on ISSA-HKELM according to claim 3, characterized in that, The dynamic weighting factor The value and the maximum number of iterations K and the current iteration number k Related, and with the current iteration number k The increase leads to a decrease; In the early stages of iteration, when the number of iterations is small... The value is larger; in the later stages of the iteration, when the number of iterations increases, The value is relatively small.

5. The short-term load forecasting method based on ISSA-HKELM according to claim 3, characterized in that, The dynamic weighting factor The calculation formula and the improved discoverer location update method are shown below: in For the previous generation j The global optimal solution of dimension. Indicates the current iteration number; ; Indicates the first In the nth iteration Only sparrows in the first The position of the dimension; Indicates the maximum number of iterations; Random numbers in the data; R 2 represents the warning value and R 2 0,1 ; ST Represents the safety threshold and ST 0.5,1 ; Q Represents a random value that follows a normal distribution; L It is a function whose internal elements are all 1. The matrix.

6. The short-term load forecasting method based on ISSA-HKELM according to claim 1, characterized in that, The root mean square error, mean absolute percentage error, mean absolute error, and good fit were selected as evaluation metrics for the ISSA-HKELM model.

Images