User-side distributed wind power multi-time scale prediction method and system and medium

Abstract

The invention provides a user side distributed wind power multi-time scale prediction method and system and a medium, and the method comprises the steps: firstly obtaining the wind power generation power, wind speed, temperature and humidity, and extracting the characteristics of data at different time scales through a convolution module; the method comprises the steps of firstly, optimizing an LSTM network structure by adopting an adam optimizer, improving an FTTA optimization algorithm, optimizing hyper-parameters of the LSTM by utilizing the improved FTTA optimization algorithm, normalizing wind power generation power, wind speed, temperature and humidity data, and finally inputting the normalized data into an LSTM model of the improved FTTA optimization algorithm for prediction. According to the method provided by the invention, prediction is carried out on multiple scales through the optimized LSTM model, so that the prediction precision of the wind power is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method, system and medium for predicting multi-time scales of distributed wind power at a user side. Background Art

[0002] As a new type of energy, wind energy is widely used due to its unlimited reserves, safety, and cleanliness, and has been vigorously developed by various countries. The non-stationarity of wind speed leads to greater randomness and volatility in wind power generation. On the other hand, the proportion of new capacity of distributed wind power on the user side has increased year by year, and the development trend of centralized and distributed wind power generation is obvious. However, since distributed wind power on the user side often lacks special meteorological monitoring devices, and a large number of newly built wind farms each year lack sufficient historical data, it is difficult to establish a high-precision prediction model, which makes distributed wind power on the user side in a "blind spot" of new energy regulation for a long time. Therefore, it is necessary to accurately predict the power of distributed wind power resources on multiple time scales to provide accurate data support for the optimization of power resource scheduling on the user side.

[0003] With the rapid development of deep learning and its significant success in various fields, wind power prediction methods based on deep learning have been widely studied, and the prediction effect is better than that of traditional statistical methods. Although wind power prediction methods are receiving more and more attention from researchers, neural networks still have problems such as slow convergence speed and low accuracy caused by parameters, so a high-precision user-side distributed wind power prediction method is urgently needed. Summary of the invention

[0004] The purpose of the present invention is to solve the deficiencies of the above-mentioned prior art and to provide a multi-time scale prediction method, system and medium for distributed wind power on the user side. The method proposed in the present invention performs prediction on multiple scales through an optimized LSTM model, thereby improving the prediction accuracy of wind power.

[0005] A multi-time scale prediction method for distributed wind power at a user side, comprising:

[0006] Step 1: Obtain data on wind power, wind speed, temperature, and humidity, and clean up outliers and missing values ​​in the data;

[0007] Step 2: Extract the features of data at different time scales through the convolution module to obtain the collected data;

[0008] Step 3: Use the adam optimizer to optimize the network structure bias parameter b and network weight parameter W of the LSTM model;

[0009] Step 4: Using an improved FTTA optimization algorithm to optimize the hyperparameters of the LSTM model to obtain an LSTM optimized prediction model, wherein the hyperparameters include a learning rate, a regularization coefficient, and the number of neurons;

[0010] Step 5: Normalize the collected data and input it into the LSTM optimization prediction model for prediction.

[0011] Furthermore, in step 2, four convolutional layers with convolution kernel sizes of 3×3, 5×5, 7×7 and 9×9 are selected and connected in parallel as a multi-scale feature extraction module; the step size of each convolutional layer is set to 2, the number of channels is set to 8, the padding mode is set to "same", and the padding value is set to 0.

[0012] Furthermore, step 3 specifically includes:

[0013] Initialize V db 、V dW , S db , S dW ;

[0014] In the tth iteration, V is calculated using the mini-batch gradient descent method db and V dW , specifically:

[0015] V db =β1V db +(1-β1)db;

[0016] V dW =β1V dW +(1-β1)dW;

[0017] Calculate S using RMSprop db and S dW , specifically:

[0018] S db =β2S db +(1-β2)db 2 ;

[0019] S dW =β2S dW +(1-β2)dW 2 ;

[0020] Among them, β1 and β2 are the attenuation coefficients of the first-order and second-order estimation matrices, and their values ​​are 0.9 and 0.999 respectively. db and V dW are the first-order estimation matrices of the network structure bias parameter b and the network weight parameter W, respectively, S db and S dWare the second-order estimation matrices of the network structure bias parameter b and the network weight parameter W, db and dW are V db and V dW The gradient value of

[0021] According to V db 、V dW , S db and S dW , calculate the bias correction and Specifically:

[0022]

[0023] Where k is the number of iterations, and To correct the first-order estimation matrix, and is the modified second-order estimation matrix;

[0024] Correction based on deviation and Update the network structure bias parameter b and network weight parameter W of the LSTM model, specifically:

[0025]

[0026] Among them, α is the learning rate and ε is 10 -8 The constant.

[0027] Furthermore, the improved FTTA optimization algorithm has the following specific process:

[0028] 4.1 Introduce Logistic-Sine-Cosine chaotic mapping to initialize the population to solve the uneven distribution problem caused by random initialization:

[0029]

[0030] F i,j =(ub-lb)×Rand+lb

[0031] f i,j =cos(π(4rf i,j (1-f i,j )+(1-r)sin(πf i,j )-0.5)),r∈[0,1]

[0032] In the formula, F represents the initialized population, f i,jRepresents the value of the i-th individual in the j-th dimension, i∈1,2…n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each individual represents a set of LSTM learning rate, regularization coefficient and number of neurons, ub and lb are the upper and lower bounds of the problem respectively, and Rand represents a random number between 0 and 1;

[0033] 4.2 Group training

[0034] The players will conduct collective training under the guidance of the coach. The coach will first let the players understand their own level through a series of tests. Then the players will develop their own collective training plan according to their own level. The players will be divided into four different types: Followers, Finders, Thinkers and Fluctuators. In each iteration, the players will randomly change their types.

[0035] 4.2.1 Followers

[0036]

[0037] in, represents the optimal individual in the current iteration, where k is the number of iterations, represents the updated individual, represents the individual before the update, and rand represents a random number between 0 and 1;

[0038] 4.2.2 Discoverer

[0039] In view of the fact that the discoverer strategy is prone to fall into local optimality, an escape strategy is introduced to improve it:

[0040]

[0041] in, represents the worst individual in the current iteration, esc represents the escape operator;

[0042] 4.2.3 Thinkers

[0043]

[0044] 4.2.4 Fluctuators

[0045]

[0046] Among them, t(k) is a random number with t distribution, and its degree of freedom is the current iteration number. As the degree of freedom increases, the probability of t distribution approaching the middle value (0) becomes higher and higher, and the distribution at both ends gradually decreases, and will become closer and closer to the normal distribution;

[0047] 4.3 Group training

[0048] 4.3.1 Optimal Learning

[0049]

[0050] in, represents the updated individuals during group training, represents the individual before updating during group training, represents the optimal individual in group training, p study represents the learning probability;

[0051] 4.3.2 Stochastic Learning

[0052]

[0053] in, represents a random individual during group training;

[0054] 4.2.3 Random Communication

[0055]

[0056] Among them, Randn is a normally distributed random number, p comm represents the probability of exchange, represents the updated random individuals during group training, represents a random individual before updating during group training;

[0057] 4.2.4 Random Error

[0058]

[0059] in, represents a random individual before updating on a random dimension during group training;

[0060] Among them, p error represents the probability of error;

[0061] 4.3 Additional personal training

[0062]

[0063] Among them, Gauss represents Gaussian operator, Cauchy represents Cauchy operator;

[0064] 4.4 Determine whether the current iteration has reached the maximum number of times. If not, continue iterating; otherwise, stop iterating and output the optimal individual, which is the optimal value of the LSTM learning rate, regularization coefficient and number of neurons.

[0065] Furthermore, step 5 specifically includes:

[0066] 5.1 Normalize the data, specifically:

[0067]

[0068] In the formula: Z represents the normalized data, z represents the data, z max Indicates the maximum value of the data, z min Indicates the minimum value of data;

[0069] 5.2 The LSTM model prediction process is as follows:

[0070] f t =σ(W f *[h t-1 , x t ]+b f );

[0071] i t =σ(W i *[h t-1 , x t ]+b i );

[0072] o t =σ(W o *[h t-1 , x t ]+b o );

[0073] Among them, h t-1 is the hidden layer state at time t-1, f t For the forget gate, i t is the input gate, o t is the output gate, x t is the input at time t, W f , W c , W i and W o is the weight coefficient matrix, b f is the bias of the forget gate, b i is the bias of the input gate, b o is the bias of the output gate, σ represents the Sigmoid activation function;

[0074] C t =f t +c t-1 +i t +tanh(W c *[h t-1 , x t ]+b c );

[0075] h t =o t *tan(c t );

[0076] In the formula, c t-1 is the state of the gated unit at time t-1, and tanh represents the hyperbolic sine activation function.

[0077] A user-side distributed wind power multi-time scale prediction system, comprising: a computer-readable storage medium and a processor;

[0078] The computer-readable storage medium is used to store executable instructions;

[0079] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the multi-time scale prediction method for user-side distributed wind power.

[0080] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the multi-time-scale prediction method for distributed wind power at the user side.

[0081] The present invention adopts convolution module to collect features at different scales, reduces the influence of invalid information and noise data on the network training process, reduces the calculation amount and memory consumption of subsequent layers, shortens the training time, uses adam optimizer and improved FTTA algorithm to optimize the network structure and hyperparameters of LSTM, and improves the prediction accuracy of wind power. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A flowchart of a method for multi-time scale prediction of distributed wind power at a user side according to an embodiment of the present invention;

[0083] Figure 2 A flowchart of an improved FTTA optimization algorithm according to an embodiment of the present invention;

[0084] Figure 3 This is a comparison diagram of the convergence curves of the improved FTTA optimization algorithm according to the embodiment of the present invention and the existing FTTA optimization algorithm. DETAILED DESCRIPTION

[0085] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] Reference Figure 1 The embodiment of the present invention provides a photovoltaic power prediction method based on similarity recombination, comprising the following steps:

[0087] Step 1: Obtain data on wind power, wind speed, temperature, and humidity, and clean up outliers and missing values;

[0088] Step 2: Extract features of different scales of data through the convolution module to obtain the collected data; specifically, in step 2, four convolution layers with convolution kernel sizes of 3×3, 5×5, 7×7 and 9×9 are selected and connected in parallel as a multi-scale feature extraction module; set the step size of each convolution layer to 2, the number of channels to 8, the padding mode to "same", and the padding value to 0.

[0089] Step 3: Use the adam optimizer to optimize the network structure bias parameter b and network weight parameter W of the LSTM model;

[0090] The adam optimizer is used to optimize the network structure bias parameter b and network weight parameter W of the LSTM model. The specific steps are as follows:

[0091] Initialize V db 、V dW , S db , S dW :

[0092] In the tth iteration, V is calculated using the mini-batch gradient descent method db and V dW , specifically:

[0093] V db =β1V db +(1-β1)db;

[0094] V dW =β1V dW +(1-β1)dW;

[0095] Calculate S using RMSprop db and S dW , specifically:

[0096] S db =β2S db +(1-β2)db 2 ;

[0097] S dW =β2S dW +(1-β2)dW 2 ;

[0098] Among them, β1 and β2 are the attenuation coefficients of the first-order and second-order estimation matrices, and their values ​​are 0.9 and 0.999 respectively. db and V dWare the first-order estimation matrices of the network structure bias parameter b and the network weight parameter W, respectively, S db and S dW are the second-order estimation matrices of the network structure bias parameter b and the network weight parameter W, db and dW are V db and V dW The gradient value of

[0099] According to V db 、V dW , S db and S dW , calculate the bias correction and Specifically:

[0100]

[0101] Where k is the number of iterations, and To correct the first-order estimation matrix, and is the modified second-order estimation matrix;

[0102] Correction based on deviation and Update the network structure bias parameter b and network weight parameter W of the LSTM model, specifically:

[0103]

[0104] Among them, α is the learning rate and ε is 10 -8 The constant.

[0105] Step 4: Use the improved FTTA optimization algorithm to optimize the hyperparameters of the LSTM model to obtain the LSTM optimized prediction model; the hyperparameters include learning rate, regularization coefficient and number of neurons; Figure 2 As shown, the specific process of the improved FTTA optimization algorithm is as follows:

[0106] 4.1 Introduce Logistic-Sine-Cosine chaotic mapping to initialize the population to solve the uneven distribution problem caused by random initialization.

[0107]

[0108] F i,j =(ub-lb)×Rand+lb

[0109] f i,j =cos(π(4rf i,j (1-f i,j )+(1-r)sin(πf i,j)-0.5)),r∈[0,1]

[0110] In the formula, F represents the initialized population, f i,j Represents the value of the i-th individual in the j-th dimension, i∈1,2…n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each individual represents a set of LSTM learning rate, regularization coefficient and number of neurons, ub and lb are the upper and lower bounds of the problem respectively, and Rand represents a random number between 0 and 1.

[0111] 4.2 Group training

[0112] The players will train collectively under the guidance of the coach, who will first let the players know their level through a series of tests (physical functions). Then the players will develop their own collective training plan based on their level. We divide the players into four different types: Followers, Finders, Thinkers, and Fluctuators. In each iteration, the players will randomly change their type.

[0113] 4.2.1 Followers

[0114]

[0115] in, represents the optimal individual in the current iteration, where k is the number of iterations, represents the updated individual, represents the individual before updating, and rand represents a random number between 0 and 1.

[0116] 4.2.2 Discoverer

[0117] Aiming at the shortcoming that the discoverer strategy is prone to fall into local optimality, an escape strategy is introduced to improve it.

[0118]

[0119] in, represents the worst individual in the current iteration, and esc represents the escape operator.

[0120] 4.2.3 Thinkers

[0121]

[0122] 4.2.3 Fluctuators

[0123]

[0124] Among them, t(k) is a random number with t distribution, and its degree of freedom is the current number of iterations. As the degree of freedom increases, the probability that the t distribution is close to the middle value (0) becomes higher and higher, and the distribution at both ends gradually decreases, and will become closer and closer to the normal distribution.

[0125] 4.3 Group training

[0126] 4.3.1 Optimal Learning

[0127]

[0128] in, represents the updated individuals during group training, represents the individual before updating during group training, represents the optimal individual in group training, p study represents the learning probability.

[0129] 4.3.2 Stochastic Learning

[0130]

[0131] in, Represents a random individual during group training.

[0132] 4.2.3 Random Communication

[0133]

[0134] Among them, Randn is a normally distributed random number, p comm represents the probability of exchange, represents the updated random individuals during group training, Represents a random individual before updating during group training.

[0135] 4.2.4 Random Error

[0136]

[0137] in, represents a random individual before updating on a random dimension during group training.

[0138] Among them, p error represents the error probability,

[0139] 4.3 Additional personal training

[0140]

[0141] Among them, Gauss represents Gaussian operator and Cauchy represents Cauchy operator

[0142] 4.4 Determine whether the current iteration has reached the maximum number of times. If not, continue iterating; otherwise, stop iterating and output the optimal individual, which is the optimal value of the LSTM learning rate, regularization coefficient and number of neurons.

[0143] Depend on Figure 3 It can be seen that the improved FTTA optimization algorithm of the present invention is superior to the existing FTTA optimization algorithm in terms of convergence accuracy and convergence speed.

[0144] Step 5: Normalize the collected data and input it into the optimized LSTM model for prediction.

[0145] 5.1 Normalize the data, specifically:

[0146]

[0147] In the formula: Z represents the normalized data, z represents the data, z max Indicates the maximum value of the data, z min Indicates the minimum value of data;

[0148] 5.2 The LSTM model prediction process is as follows:

[0149] f t =σ(W f *[h t-1 , x t ]+b f );

[0150] i t =σ(W i *[h t-1 , x t ]+b i );

[0151] o t =σ(W o *[h t-1 , x t ]+b o );

[0152] Among them, h t-1 is the hidden layer state at time t-1, f t For the forget gate, i t is the input gate, o t is the output gate, x t is the input at time t, W f , W c , W i and W o is the weight coefficient matrix, b f is the bias of the forget gate, b i is the bias of the input gate, bo is the bias of the output gate, σ represents the Sigmoid activation function;

[0153] C t =f t +c t-1 +i t +tanh(W c *[h t-1 , x t ]+b c );

[0154] h t =o t *tan(c t );

[0155] In the formula, c t-1 is the state of the gated unit at time t-1, and tanh represents the hyperbolic sine activation function.

[0156] Another aspect of the present invention provides a user-side distributed wind power multi-time scale prediction system, comprising: a computer-readable storage medium and a processor;

[0157] The computer-readable storage medium is used to store executable instructions;

[0158] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the multi-time scale prediction method for user-side distributed wind power.

[0159] Another aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-time-scale prediction method for distributed wind power at the user side described in the first aspect.

[0160] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0161] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0164] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-time scale prediction method for distributed wind power at the user side, characterized in that: The following steps are involved: Step 1: Obtain data on wind power, wind speed, temperature, and humidity, and clean up outliers and missing values ​​in the data; Step 2: Extract the features of data at different time scales through the convolution module to obtain the collected data; Step 3: Use the adam optimizer to optimize the network structure bias parameter b and network weight parameter W of the LSTM model; Step 4: Using an improved FTTA optimization algorithm to optimize the hyperparameters of the LSTM model to obtain an LSTM optimized prediction model, wherein the hyperparameters include a learning rate, a regularization coefficient, and the number of neurons; Step 5: Normalize the collected data and input it into the LSTM optimization prediction model for prediction.

2. The multi-time scale prediction method for distributed wind power at the user side according to claim 1 is characterized in that: In step 2, four convolutional layers with convolution kernel sizes of 3×3, 5×5, 7×7, and 9×9 are selected and connected in parallel as a multi-scale feature extraction module; Set the stride of each convolutional layer to 2, the number of channels to 8, the padding mode to "same", and the padding value to 0.

3. The multi-time scale prediction method for distributed wind power at the user side according to claim 1 is characterized in that: Step 3 specifically includes: Initialize V db 、V dW , S db , S dW ; In the tth iteration, V is calculated using the mini-batch gradient descent method db and V dW , specifically: V db =β1V db +(1-β1)db; V dW =β1V dW +(1-β1)dW; Calculate S using RMSprop db and S dW , specifically: S db =β2S db +(1-β2)db 2 ; S dW =β2S dW +(1-β2)dW 2 ; Among them, β1 and β2 are the attenuation coefficients of the first-order and second-order estimation matrices, and their values ​​are 0.9 and 0.999 respectively. db and V dW are the first-order estimation matrices of the network structure bias parameter b and the network weight parameter W, respectively, S db and S dW are the second-order estimation matrices of the network structure bias parameter b and the network weight parameter W, db and dW are V db and V dW The gradient value of According to V db 、V dW , S db and S dW , calculate the bias correction and Specifically: Where k is the number of iterations, and To correct the first-order estimation matrix, and is the modified second-order estimation matrix; Correction based on deviation and Update the network structure bias parameter b and network weight parameter W of the LSTM model, specifically: Among them, α is the learning rate and ε is 10 -8 The constant.

4. The multi-time scale prediction method for distributed wind power at the user side according to claim 1 is characterized in that: The specific process of the improved FTTA optimization algorithm is as follows: 4.1 Introduce Logistic-Sine-Cosine chaotic mapping to initialize the population to solve the uneven distribution problem caused by random initialization: F i,j =(ub-lb)×Rand+lb f i,j =cos(π(4rf i,j (1-f i,j )+(1-r)sin(πf i,j )-0.5)),r∈[0,1] In the formula, F represents the initialized population, f i,j Represents the value of the i-th individual in the j-th dimension, i∈1,2…n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; each individual represents a set of LSTM learning rate, regularization coefficient and number of neurons, ub and lb are the upper and lower bounds of the problem respectively, and Rand represents a random number between 0 and 1; 4.2 Group training The players will conduct collective training under the guidance of the coach. The coach will first let the players understand their own level through a series of tests. Then the players will develop their own collective training plan according to their own level. The players will be divided into four different types: Followers, Finders, Thinkers and Fluctuators. In each iteration, the players will randomly change their types. 4.2.1 Followers in, represents the optimal individual in the current iteration, where k is the number of iterations, represents the updated individual, represents the individual before the update, and rand represents a random number between 0 and 1; 4.2.2 Discoverer In view of the fact that the discoverer strategy is prone to fall into local optimality, an escape strategy is introduced to improve it: in, represents the worst individual in the current iteration, esc represents the escape operator; 4.2.3 Thinkers 4.2.4 Fluctuators Among them, t(k) is a random number with t distribution, and its degree of freedom is the current iteration number. As the degree of freedom increases, the probability of t distribution approaching the middle value (0) becomes higher and higher, and the distribution at both ends gradually decreases, and will become closer and closer to the normal distribution; 4.3 Group training 4.3.1 Optimal Learning in, represents the updated individuals during group training, represents the individual before updating during group training, represents the optimal individual in group training, p study represents the learning probability; 4.3.2 Stochastic Learning in, represents a random individual during group training; 4.2.3 Random Communication Among them, Randn is a normally distributed random number, p comm represents the probability of exchange, represents the updated random individuals during group training, represents a random individual before updating during group training; 4.2.4 Random Error in, represents a random individual before updating on a random dimension during group training; Among them, p error represents the error probability; 4.3 Additional personal training Among them, Gauss represents Gaussian operator, Cauchy represents Cauchy operator; 4.4 Determine whether the current iteration has reached the maximum number of times. If not, continue iterating; otherwise, stop iterating and output the optimal individual, which is the optimal value of the learning rate, regularization coefficient and number of neurons of LSTM.

5. The multi-time scale prediction method for distributed wind power at the user side according to claim 1 is characterized in that: Step 5 specifically includes: 5.1 Normalize the data, specifically: In the formula: Z represents the normalized data, z represents the data, z max Indicates the maximum value of the data, z min Indicates the minimum value of data; 5.2 The LSTM model prediction process is as follows: f t =σ(W f *[h t-1 ,x t ]+b f ); i t =σ(W i *[h t-1 ,x t ]+b i ); the t =σ(W o *[h t-1 ,x t ]+b o ); Among them, h t-1 is the hidden layer state at time t-1, f t For the forget gate, i t is the input gate, o t is the output gate, x t is the input at time t, W f , W c , W i and W o is the weight coefficient matrix, b f is the bias of the forget gate, b i is the bias of the input gate, b o is the bias of the output gate, σ represents the Sigmoid activation function; c t =f t +c t-1 +i t +tanh(W c *[h t-1 ,x t ]+b c ); h t =o t *tan(c t ); In the formula, c t-1 is the state of the gated unit at time t-1, and tanh represents the hyperbolic sine activation function.

6. A multi-time scale prediction system for distributed wind power at the user side, comprising: A computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the user-side distributed wind power multi-time scale prediction method according to any one of claims 1-5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for multi-time-scale prediction of user-side distributed wind power according to any one of claims 1 to 5 is implemented.

Images