TW201635224A - Method of short-term wind power generation forecasting - Google Patents

Abstract

An intelligent method of forecasting short-term wind power generation includes providing a persistence algorithm to forecast wind speed and wind power generation in the future according to the wind speed and wind power generation in the present; providing a radial basis function neural network algorithm to forecast wind power generation in the future; providing a back propagation neural network algorithm to forecast wind power generation in the future; providing a smart mathematical forecasting model; and providing an artificial bee colony algorithm to solve weighting coefficients of the smart mathematical forecasting model to obtain the optimal wind power generation forecasting value.

Description

Intelligent short-term power generation forecasting method

The invention relates to a short-term power generation quantity prediction method, in particular to a smart short-term power generation quantity prediction method which uses an artificial bee colony algorithm to solve the weight coefficient of the optimized compound prediction method.

As petrochemical energy is declining and oil prices are rising, it is an indisputable fact that energy supply is becoming shorter. In addition, as environmental issues have gradually gained attention, after the implementation of the Kyoto Protocol in 2005, the total emissions of carbon dioxide in the country have been regulated, thus limiting the use of fossil fuels. In the case that the traditional heat generating units are not easy to add, new alternative energy development is a necessary solution. Therefore, all advanced countries do not aim at the development of renewable energy application technology, and wind power generation is one of the most valuable renewable energy sources. Especially in Taiwan, where electricity consumption is increasing, the development of wind power technology is urgent. of. In recent years, wind turbines have been installed on the western coast of Taiwan. The main reason is that wind power has high efficiency, high power density and low pollution.

Although wind power has many of the above advantages, there is still a technical problem to be overcome in application, which is the prediction of wind power generation. Because the wind turbine's power generation will change with meteorological conditions and wind speed, unpredictable changes in wind turbine power generation will lead to an increase in power system backup capacity and an increase in operating costs. Power system dispatchers must predict changes in wind turbine generation to schedule the reserve capacity and manage the system. Accurate wind power forecasting helps to improve the design of liberalized market mechanisms, the immediate operation control of the grid, the development of grid-connected standards, and the improvement of power quality. Since wind power has become an important way to develop renewable energy, the development of related technologies for wind power forecasting has become an extremely important part of power engineering.

Therefore, how to design a smart short-term power generation forecasting method, provide a large elasticity, can be applied to various types of wind turbine generation generation forecast, and use artificial bee colony algorithm to solve the weight coefficient of each individual prediction method, Effectively improve the prediction accuracy of the intelligent forecasting method, and effectively and correctly predict the power generation of the wind power generation system in different seasons, and provide an important tool for the power system dispatcher to control the operation of the relevant power system to improve the power system. The purpose of stability and lowering the operating cost of the system is a major issue that the inventors of the present invention have tried to overcome and solve.

It is an object of the present invention to provide a smart short-term power generation amount prediction method that overcomes the problems of the prior art. The prediction method includes: (a) providing a continuous method to predict the predicted wind speed and wind power generation at a future time point based on the wind speed and wind power generation measured at the current time point: , ;among them, For the current time, ( ) for the future time, For the current wind speed, For future wind speed, For current wind power generation as well (b) provide a radial basis function-like neural network algorithm that predicts future wind power generation using a network architecture that includes an input layer, a hidden layer, and an output layer; (c) provides a An inverse transfer-like neural network algorithm predicts future wind power generation using a network architecture including an input layer, a hidden layer, and an output layer; (d) according to the persistence method, the radial basis function-like neural network algorithm And the inverse transfer neural network algorithm provides a smart mathematical prediction model ;among them, , ;as well as ( t = 1, 2, ..., L ) is the actual power generation time series data, M is the number of individual prediction methods, and L is the number of samples. ( i = 1, 2, ..., M , t = 1, 2, ..., L ) is the predicted value of the i-th prediction method, = - For prediction errors, The weighting factor for the i-th prediction method, for Estimated value, and a predictive value for the intelligent predictive method; and (e) providing an artificial bee colony algorithm to solve the weighting coefficient of the intelligent mathematical predictive model To obtain the best value of wind power generation; .

In order to further understand the technology, the means and the effect of the present invention in order to achieve the intended purpose, refer to the following detailed description of the invention and the accompanying drawings. The detailed description is to be understood as illustrative and not restrictive.

The technical content and detailed description of the present invention are as follows:

Please refer to FIG. 1 , which is a system architecture diagram of a power generation quantity prediction system for a smart short-term wind power generation system of the present invention. The construction of the power generation quantity prediction system first establishes a training database from the power generation data of the actual wind power generation system, and then establishes a short-term wind power generation quantity prediction system through the training program. The intelligent short-term wind power generation system power generation quantity prediction system combines three short-term wind power generation system power generation quantity prediction systems: (1) continuous method, (2) radial basis function-like neural network algorithm, and (3) reverse-transfer-like nerve In the network algorithm, the input signal is first predicted by three short-term wind power generation system prediction systems, and the predicted value of the starting electricity is calculated by the equation of the composite prediction method. The short-term wind power generation forecasting system inputs the data of the first 20 minutes of power generation, the first 10 minutes of power generation, the current power generation amount, and the wind speed prediction value, and uses an artificial bee colony algorithm to solve the wind power generation amount. Obtain a predicted value of power generation after 10 minutes.

The combined theoretical basis of intelligent predictive method is to solve specific prediction problems by several different individual prediction methods. Each individual prediction method has a specific weight coefficient, and the intelligent prediction prediction value is the weight of these individual prediction methods. (weighting sum). The mathematical model of the intelligent prediction method can be expressed as follows:

(1)

(2)

among them, ( t = 1, 2, ..., L ) is the actual power generation time series data, M is the number of individual prediction methods, and L is the number of samples. ( i = 1, 2, ..., M , t = 1, 2, ..., L ) is the predicted value of the i-th prediction method, = - For prediction errors, The weighting factor for the i-th prediction method, for Estimated value, and The predicted value for the intelligent forecasting method.

In the structure of the intelligent prediction method, the determination of the weight coefficient in each individual prediction method is the most important step, and the weight coefficient can be realized by solving the optimization problem of the absolute error minimum value of the composite prediction method. The mathematical model of the optimization problem can be expressed as follows:

(3)

Hereinafter, the algorithm employed in the present invention will be described in detail.

1. Artificial bee colony (ABC)

The invention will solve the (3) formula by the artificial bee colony algorithm, and the artificial bee colony algorithm is based on the bee group foraging behavior mode, and the optimization algorithm is developed, and the calculation process of the bio-group wisdom algorithm is The development program and the exploration mechanism feature not only avoid the solution process falling into the local solution, but also improve the computational efficiency of solving the optimization problem. The invention will determine the weight coefficient combination of the three individual prediction methods by the artificial bee colony algorithm, and then predict the power generation of the short-term wind power generation system by the formula (2).

According to the process of bee colony feeding, worker bees can be divided into three types of characters: collecting bees, standby bees and scout bees. In the early stage of foraging, because the worker bee group does not have any food source information, the worker bees randomly search for the role of the scout bee. Once the food source is found, the bee is collected to collect the nectar of the food source until the food source After the nectar is collected, it is converted into the role of the standby bee or the role of the scout bee, so that the subsequent honey collecting work can be carried out. When the worker bees give up the food source, if they are converted into the role of the standby bee, they fly back to the hive to rest and wait for the next scout to share the food source message; otherwise, if they are converted into the role of the scout bee, they continue to randomly search for new ones. Food source location.

The artificial bee colony algorithm is used to solve the weight coefficient combination problem of the individual prediction methods. The food source position is the weight coefficient combination of the individual prediction methods, and the nectar quantity of the food source can be regarded as the target function value of the optimization problem. That is, the absolute error of the composite prediction method. Before the start of the artificial bee colony algorithm, it is necessary to input the relevant information of the individual prediction methods. In the calculation process of the artificial bee colony algorithm, the actual worker bee is searched for the behavior pattern of the nectar, which can be divided into a development program and an exploration program. And if this program is combined and run, not only can the local search work be performed, but also the search for the global solution.

The artificial bee colony algorithm randomly generates N initial food source positions and N collecting bees, and each food source is mapped to a collecting bee. The position of the food source is the self-changing parameter for solving the optimization problem. In the invention, the weight coefficient combination of the individual prediction methods, and the amount of nectar contained in each food source is the target function value for solving the optimization problem, which is the absolute error of the composite prediction method in the present invention. Once the food source location is initialized, the development and exploration programs can be executed.

During the actual bee colony foraging, when the amount of nectar in the food source is insufficient or has been collected, the collecting bee will be turned into a scout bee to find a new food source location. In the artificial bee colony algorithm, the first step is to set the number of nectar of the food source to collect whether the nectar of the food source is sufficient, that is, the target function value (absolute error of the composite prediction method), through the setting procedure. And calculus, if the target function 无法 value can not be improved, the group solution is abandoned, and another set of new solutions is randomly generated to avoid premature convergence or partial solution to the calculation process. The focus of the artificial bee colony algorithm is to collect bees to explore food sources rich in nectar and to share this food information with other worker bees, in order to find a more abundant source of nectar in the vicinity of the food source, thereby enhancing regional search. Ability, if the amount of nectar in the food source is insufficient (the target function cannot be reduced), the food source is abandoned and the scout is randomly searched for new food sources to enhance the calculation of new feasible solutions.

When the artificial bee colony algorithm performs the development process, the standby bee develops a new food source, and the initial food source location is updated to develop the remaining nectar-rich food source while avoiding the exhaustion of the original food source nectar. If the standby bee finds a new source of food, compare the target value of the new food source with the original food source position. If the new food source has a better target value than the original position, replace the old food source with the new food source. When performing the exploration process, the investigation bee will be called to visit various places, and random search for possible food sources (new feasible solutions), planning the behavior of the investigation bee to explore new food sources, and using the detection bee as the basis for decision-making, eliminating the target function value Poor food sources and replaced with new food sources. Until the iterative generation reaches the upper limit of the iteration set by the program, the artificial bee colony algorithm calculus is stopped and the optimal solution is output. In the present invention, the optimal solution is the optimal weight coefficient combination of the individual prediction methods.

The invention applies the group optimization characteristic of the artificial bee colony algorithm to solve the weight coefficient combination of the individual prediction methods, and the basic concept is that the weight coefficient combination of the individual prediction methods is set as the artificial bee colony algorithm for each food source, through the bee colony. The optimal solution is found in the optimization process of the search space to determine the optimal value of each weight coefficient combination of the individual prediction methods (as shown in steps S108 and S110 of FIG. 2). The execution steps are detailed later.

Step 1: Set the initial data. In this step, set the number of food sources, variable dimensions, upper and lower limits of design variables, upper limit of iterations, and food source restrictions. Generate N sets of variable combinations by (4) and calculate variables with (5) Adapt to the value and record the best solution.

(4)

among them, J-th combination for the i-th variable variable i = 1, 2, ..., N, N is the number of food sources, also referred to as the number of groups, i.e. the number of combinations of variables; j = 1, 2, ... , D, D is a dimension; 𝛼 is a random random number between 0 and 1; versus The upper and lower limits of the design variables for the variable combination.

The above adaptation values are calculated as shown in equation (5):

(5)

among them, for The cost value is also the sum of the weights of the absolute errors of the individual prediction methods.

Step 2: Generate new variables by neighborhood search. In this step, generate new design variables with (6) And the limit cannot exceed the upper and lower limits.

(6)

among them, The ith new design variable for the jth variable combination, The i-th old design variable for the j- th variable combination, For a randomly generated value between -1 and 1, a combination of variables randomly selected for the population, .

Step 3: Decide whether to update the variable combination. Calculated variable combination Adaptation value, if Preferably, the adaptation value is updated to change the combination of variables.

Step 4: Decide whether to enter the observation bee stage:

Variable combination Randomly generate a random number x between 0 and 1, and calculate the combination of variables Screening probability, if x is less than the screening probability, the variable combination Enter the observation bee stage.

The above screening probability is calculated as shown in (7):

(7)

among them, Combination of variables Screening probability, for The fitness value, N is the number of food sources.

Step 5: Observe the bee stage. In this step, generate new design variables with (6) And the limit cannot exceed the upper and lower limits. Calculated variable combination For the fitness value, select a combination of variables with better fitness values.

Step 6: Decide whether to enter the scout bee stage. In this step, it is judged whether there is a food source reaching the restriction condition: if the restriction condition is reached, the food source is abandoned to enter the detection bee stage, otherwise the detection bee stage is skipped.

Step 7: Detect the bee stage. Generating a new combination of variables with the formula (4) for the j-th set of variables , calculate the fitness value.

Step 8: Check if the end condition is met. If the search iteration number reaches the set top generation limit value, the iteration is ended and the optimal weight coefficient combination is output; otherwise, return to step 2 to continue the iteration.

2, persistence method (persistence method)

The continuous method is not only a simple principle, but also the most economical method of wind speed or power prediction. Power companies in various countries often use the continuous method as the benchmark method for predicting ultra-short-term wind power generation. The principle of the continuation method is based on a simple assumption: when the predicted time interval is extremely short, the wind speed or wind power generation at the current measurement time point and the future prediction time point will be the same. Under certain conditions, the continuous method is more accurate for the prediction of ultra-short-term wind power than other wind power generation prediction methods (as shown in step S102 of Fig. 2).

The principle of the continuation method is based on a simple assumption: if the current time point The measured wind speed and wind power generation are versus , then the future time point ( The predicted wind speed and wind power generation are:

(8)

(9)

3. Radial basis function (RCF)-like neural network algorithm

Neural networks are often used to solve the problem of parameter estimation. RBF-like neural networks have the advantages of high robustness and strong approximation in various neural networks. Therefore, they are particularly suitable for nonlinear estimation. The short-term wind power prediction system will also use the RBF-like neural network as one of three individual prediction systems (as shown in step S104 of FIG. 2). The basic principle of RBF neural network comes from function approximation. The basic architecture of RBF neural network is shown in Figure 3. It is a three-layer network architecture composed of input layer, hidden layer and output layer.

If the input layer dimension of an RBF-like neural network is d , the hidden layer dimension is q , and the output layer dimension is m , the RBF-like neural network is a d -dimensional to m -dimensional mapping relationship, as shown in the following equation:

(10)

Where the input vector is ={ , for i = 1, 2, ..., d}, the output vector is ={ , for i = 1, 2, ..., m}.

Radial basis function of each node of the hidden layer of the RBF-like neural network Usually it is a Gaussian function, as shown in the following equation:

, for j = 1, 2, ..., q (11)

among them, For the width of the Gaussian function, Is the center of the Gaussian function.

Output value of each node of the output layer of the RBF-like neural network , as shown below:

, for k = 1, 2, ..., m (12)

among them, To hide the weight between the jth node of the layer and the kth node of the output layer, The output value of the jth node of the hidden layer.

According to the RBF-like neural network architecture, when the training data pair is established, the number of input and output neurons is determined. As for the number of hidden layer neurons, it is usually regarded as a subset of the input data, that is, the number of neurons is determined by the probability density function of the input data. When the number of hidden layer neurons is too large, the network learning effect is better, but the training time is bound to increase, so a suitable number of neurons will help to improve the network learning effect. In the present invention, the orthogonal least squares theory is used to select the most. The number of good hidden layer nodes.

4. Back propagation (BP) neural network algorithm

In all kinds of neural networks, BP neural networks have the advantages of easy convergence and corresponding mapping, so they are particularly suitable for prediction. The short-term wind power prediction system of the present invention will also adopt BP neural networks for three types. One of the short-term prediction systems (as shown in step S106 of Fig. 2). The basic architecture of the BP-like neural network is shown in Figure 4. It is a three-layer network architecture consisting of an input layer, a hidden layer, and an output layer.

If the input layer dimension of a BP neural network is d , the hidden layer dimension is q , and the output layer dimension is m , the BP neural network is a d -dimensional to m -dimensional mapping relationship, as shown in (13):

(13)

Where the input vector is ={ , for i = 1, 2, ..., d}, the output vector is ={ , for i = 1, 2, ..., m}.

Input value of the jth node of the hidden layer of the BP-like neural network As shown in (14):

(14)

among them, The i-th input layer nodes to the right between the j-th hidden layer node weight, To hide the partial weight of the jth node of the layer.

The output value of the jth node of the hidden layer of the BP-like neural network , as shown in (15):

(15)

among them, To hide the input value of the jth node of the layer.

Input value of the jth node of the output layer of the BP-like neural network As shown in (16):

(16)

among them, To hide the weight between the i- th node of the layer and the j- th node of the output layer, The offset value of the jth node of the output layer.

The output value of the jth node of the output layer of the BP-like neural network As shown in (17):

(17)

among them, The input value of the jth node of the output layer.

According to the BP-like neural network architecture, when the training data pair is established, the number of input and output neurons is determined. As for the number of hidden layer neurons, it is usually regarded as a subset of the input data, that is, the number of neurons is determined by the probability density function of the input data. When the number of hidden layer neurons is too large, the network learning effect is better, but the training time is bound to increase, so a suitable number of neurons will help to improve the network learning effect. In the present invention, the orthogonal least squares theory is used to select the most. The number of good hidden layer nodes.

The BP-like neural network architecture adopted by the present invention is divided into three layers as shown in FIG. 4, wherein the first input layer has four nodes, respectively, the actual actual power generation of the wind turbine and the actual power generation in the first 10 minutes. 4 kinds of input signals such as the actual power generation and wind speed prediction in the first 20 minutes; the second layer is the hidden layer, and the number of nodes will be adjusted according to the optimal number of hidden layer nodes selected by the orthogonal least squares theory; the third layer is In the output layer, only one node is the predicted value of wind power generation after 10 minutes.

In addition, the orthogonal least squares (OLS) method used in the present invention determines the number of optimal hidden layer nodes of the neural network to be added. The orthogonal least squares method mainly uses Gram-Schmidt orthogonal theory to decompose an arbitrary vector matrix into a set of orthogonal basis vectors, and makes the space covered by the same as the space covered by the original vector matrix, so if this theory is applied to the selection of RBF The number of hidden layer nodes of the neural network and the BP neural network is sufficient to cover the space covered by the original training data. Assuming Y is the expected output vector of the training data, then:

(18)

Or written as a matrix type

(19)

Wherein, ^1, m is the number of neurons in the output layer, ^{Y Î Â m '1, B} Î Â m' q, E Î Â m '1, WÎ Â q'.

According to the Gram-Schmidt orthogonal theory, the B matrix is decomposed into a set of orthogonal basis vectors as follows:

(20)

Wherein, A Î Â ^{q 'q} is an upper triangular matrix, D Î Â ^{1' q} matrix is an orthogonal basis vectors. And

(twenty one)

among them, An element of the positive diagonal matrix H. Integration (7) to (9) can be obtained:

(twenty two)

Where G = AW .

Since D is an orthogonal basis vector matrix, the sum of squares of the expected output values Y can be expressed as:

(twenty three)

Therefore, for the k-th node, the error rate is lowered:

(twenty four)

The node's screening process is to gradually pick out the maximum error reduction rate from the training data. Point as a newly joined node and Accumulate until the set correct rate is met. In addition, in the process of selecting the number of central nodes, the present invention will adopt a Gaussian function with the same extension, that is, the standard deviation is a fixed value, as shown in the following formula:

(25)

among them, The maximum distance value between all node centers, q is the number of nodes.

In addition, in order to carry out the training and testing procedures of the intelligent wind power generation quantity prediction system, the present invention will use the one-year actual power generation data of the wind power generation unit to be tested, and the power generation data of the wind power generation unit is measured every ten minutes, therefore, one The actual power generation data of the year has a total of 52,560 data. The present invention establishes a database of the aforementioned conventional power generation data. Due to the different wind changes in the western coast of Taiwan throughout the year, the present invention will separately design a wind power generation forecasting system for the four seasons. Each quarter of the actual power generation data is built into a database, each database has 13,140 data files, each data The database selects the data files for the first two months of each season as the training database. The data files for each month after each season are filed as test data bases. These parameter databases will provide training and training for intelligent wind power generation forecasting systems. Test and other use.

In summary, the present invention has the following features and advantages:

1. The purpose of the research of the present invention is to propose a set of artificial bee colony algorithm combined with continuous method, radial basis function (RBF) neural network algorithm and reverse transfer (BP) neural network algorithm for wind power generation prediction. Intelligent short-term wind power generation forecasting method, the method is characterized by great flexibility and can be applied to various types of wind turbine generating power generation prediction;

2. Calculate the predicted value of the starting electricity quantity by the compound equation, and the weighting coefficient of each individual prediction method of the compound equation will be solved by the artificial bee colony algorithm, and the optimized intelligent prediction equation is obtained; not only the solution process is prevented from falling into the local solution. The computational efficiency of solving the optimization problem can be improved, and the prediction accuracy of the intelligent prediction method can be effectively improved;

3. The intelligent prediction method is based on the PB neural network and the RBF-like neural network. The neural network is often used to deal with nonlinear and complex classification or prediction problems. The training-like neural network can extract variables. Non-linearity, and the accuracy of prediction is higher than traditional statistical methods;

4. In the intelligent prediction method, the number of hidden layer nodes related to the RBF neural network and the PB neural network is an important parameter for determining the network prediction effect, and the present invention applies the orthogonal least square (Orthogonal Least-Squares) theory. Selecting the optimal number of hidden layer nodes can improve the prediction accuracy of RBF-like neural networks and PB-like neural networks. 5. The present invention uses the actual historical power generation data of wind turbines to perform RBF-like neural networks and PB-like nerves. Network training, all types of wind turbines only need to archive the actual historical power generation data, and then train the neural network to construct a short-term wind power prediction system dedicated to the unit. Therefore, the intelligent wind power proposed by the present invention The power generation forecasting method has great flexibility and can be applied to the short-term power generation forecast of various types of wind turbines; and 6. The short-term wind power generation system power generation forecasting system can effectively and correctly predict the wind power generation system in different seasons. Quantitatively, providing power system dispatchers with important tools for controlling the operation of relevant power systems to achieve increased power The stability of the force system and the purpose of reducing the operating cost of the system.

However, the above description is only for the detailed description and the drawings of the preferred embodiments of the present invention, and the present invention is not limited thereto, and is not intended to limit the present invention. The scope of the patent application is intended to be included in the scope of the present invention, and any one skilled in the art can readily appreciate it in the field of the present invention. Variations or modifications may be covered by the patents in this case below.

﹝this invention﹞

S102~S110‧‧‧Steps

Input layer node of x1~xd‧‧‧radial basis function neural network algorithm

B1~bq‧‧‧ hidden layer nodes of radial basis function neural network algorithm

Y1~ym‧‧‧Output layer nodes of radial basis function neural network algorithm

Input layer node of x1~xd‧‧‧ inverse transfer neural network algorithm

B1~bq‧‧‧ hidden layer nodes of inverse transfer neural network algorithm

The first figure is a system architecture diagram of the power generation quantity prediction system of the smart short-term wind power generation system of the present invention;

The second figure is a flow chart of the method for predicting the smart short-term power generation amount of the present invention;

The third figure is an architectural diagram of the radial basis function neural network algorithm of the present invention; and

The fourth figure is an architectural diagram of the inverse transfer neural network algorithm of the present invention.

S102~S110‧‧‧Steps

Claims (10)

A smart short-term power generation forecasting method includes: (a) providing a continuous method for predicting predicted wind speed and wind power generation at a future time point based on wind speed and wind power generation measured at current time points: , ;among them, For the current time, ( ) for the future time, For the current wind speed, For future wind speed, For current wind power generation as well (b) provide a radial basis function-like neural network algorithm that predicts future wind power generation using a network architecture that includes an input layer, a hidden layer, and an output layer; (c) provides one An inverse transfer-like neural network algorithm predicts future wind power generation using a network architecture including an input layer, a hidden layer, and an output layer; (d) according to the persistence method, the radial basis function-like neural network algorithm And the inverse transfer neural network algorithm provides a smart mathematical prediction model ;among them, , ;as well as ( t = 1, 2, ..., L ) is the actual power generation time series data, M is the number of individual prediction methods, and L is the number of samples. ( i = 1, 2, ..., M , t = 1, 2, ..., L ) is the predicted value of the i-th prediction method, = - For prediction errors, The weighting factor for the i-th prediction method, for Estimated value, and a predictive value for the intelligent predictive method; and (e) providing an artificial bee colony algorithm to solve the weighting coefficient of the intelligent mathematical predictive model To obtain the best value of wind power generation; . For example, the intelligent short-term power generation prediction method according to the first application of the patent scope, wherein in step (b), the RBF Neural Network is: Where d is the input layer dimension, q is the hidden layer dimension, m is the output layer dimension; and the input vector is ={ , for i = 1, 2, ..., d}, the output vector is ={ , for i = 1, 2, ..., m}. For example, the smart short-term power generation prediction method according to item 2 of the patent application scope, wherein the radial basis function of each node of the hidden layer is , is a Gaussian function: , for j = 1, 2, ..., q, where, For the width of the Gaussian function, It is the center of the Gaussian function; wherein the number of nodes of the hidden layer is determined by an orthogonal least squares theory. For example, the smart short-term power generation prediction method according to item 2 of the patent application scope, wherein the output values of the nodes of the output layer are : , for k = 1, 2, ..., m, where, To hide the weight between the jth node of the layer and the kth node of the output layer, The output value of the jth node of the hidden layer. For example, the smart short-term power generation prediction method according to the first application of the patent scope, wherein in step (c), the BP Neural Network algorithm is: Where d is the input layer dimension, q is the hidden layer dimension, m is the output layer dimension; and the input vector is ={ , for i = 1, 2, ..., d}, the output vector is ={ , for i = 1, 2, ..., m}. For example, the intelligent short-term power generation amount prediction method in the fifth application patent scope, wherein the input values of the nodes of the hidden layer are for: , among them, The i-th input layer nodes to the right between the j-th hidden layer node weight, Partial weight of the j th hidden layer node; wherein an output value of the j-th hidden layer node for: , among them, The input value of the jth node of the hidden layer; wherein the number of nodes of the hidden layer is determined by an orthogonal least squares theory. For example, the intelligent short-term power generation amount prediction method in the fifth application patent scope, wherein the input value of the j- th node of the output layer for: , among them, The system hidden layer node to the i-th output layer weights between the weight of the j-th node, Partial weight based on the j-th output layer node; wherein the output value of the output layer nodes j for: , among them, The input value of the jth node of the output layer. For example, the method for predicting the smart short-term power generation amount in the first application of the patent scope includes, in step (e), further including: (e1) setting initial data, ;among them, The i- th variable i = 1, 2, ..., N , N for the j- th variable is the number of food sources; j = 1, 2, ..., D , D are dimensions; 𝛼 is a Random random number between ~1; versus The upper and lower limits of the design variables for the combination of variables; (e2) the generation of new variables by neighborhood search, ;among them, The ith new design variable for the jth variable combination, The i-th old design variable for the j- th variable combination, For a randomly generated value between -1 and 1, a combination of variables randomly selected for the population, (e3) decide whether to update the variable combination; (e4) decide whether to enter the observation bee stage, ;among them, Combination of variables Screening probability, for Adaptation value, N is the number of food sources; (e5) observation bee stage; (e6) decide whether to enter the detection bee stage; (e7) detection bee stage; (e8) check whether the end condition is met, if the search iteration reaches the set If the upper limit of the iteration is exceeded, the iteration is ended and the optimal weight coefficient combination is output, otherwise it returns to step (e2) to continue the iteration. For example, the smart short-term power generation forecasting method in the first application scope of the patent scope, wherein before step (a), further comprises: (a01) wind power generation database archive; (a02) wind power generation database normalization; And (a03) training database file. For example, the intelligent short-term power generation quantity prediction method in the first application scope of the patent scope, wherein after step (e), further comprises: (f) anti-normalization of wind power generation prediction.