New energy power interval prediction method for mining correlation between seasonal change characteristics and meteorological characteristics

Abstract

The invention discloses a new energy power interval prediction method for mining correlation between seasonal change characteristics and meteorological characteristics. The method comprises the following steps: 1) obtaining original photovoltaic power data; 2) decomposing the original photovoltaic power data by using a moving average algorithm to obtain a trend component and a seasonal component; 3) acquiring meteorological data, and constructing a trend component prediction model and a seasonal component prediction model; 4) inputting the meteorological data and the seasonal component into the seasonal component prediction model to obtain a seasonal prediction result; 5) inputting the trend component into the trend component prediction model to obtain a trend prediction result; and 6) calculating the sum of the seasonal prediction result and the trend prediction result to obtain a final photovoltaic prediction result. According to the invention, weather information is allowed to influence seasonal components of photovoltaic power generation data without influencing trend components. The invention further provides a seasonal component prediction unit with a double-layer hierarchical attention mechanism, and attention to the relation among meteorological features, key time nodes and seasonal components is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and particularly to a new energy power interval prediction method for exploring the correlation between seasonal change characteristics and meteorological characteristics. Background Art

[0002] The real-time balance of the power system is crucial for maintaining the grid frequency. When the load of the demand side suddenly increases, the grid frequency will decrease. If not handled in time, it may trigger a chain of failures and lead to the collapse of the power grid. Weather fluctuations will cause changes in the power generation of renewable energy, further affecting the supply-demand balance. With the continuous increase in the share of renewable energy, especially photovoltaic power generation, the ability to accurately predict its output becomes increasingly important. Deterministic prediction can provide a clear picture of the future expected output, which helps to optimize the daily power generation plan. On the other hand, probabilistic prediction, especially interval prediction, is crucial for managing the inherent uncertainties in renewable energy production. These predictions are very valuable for planning the spinning reserve capacity of power plants, and the spinning reserve capacity is crucial for maintaining the stability of the power grid and ensuring the efficient operation of the power system.

[0003] The current research on photovoltaic power prediction methods has experienced an evolution process from traditional statistical models to machine learning models and then to deep learning models. The technological progress at each stage has significantly improved the ability of photovoltaic power prediction technology to handle complex non-linear relationships. Statistical models such as autoregressive models and regression analysis models were widely adopted in the early photovoltaic prediction research due to their simplicity, efficiency, and easy interpretability. These models can accurately model relatively stable and linear relationships. However, statistical models usually have difficulty in dealing with the non-linear dynamics and complex dependencies existing in photovoltaic power generation, especially when the weather conditions change rapidly or the historical patterns are insufficient to capture future fluctuations, the prediction accuracy is significantly reduced.

[0004] In recent years, the application of deep learning technology in photovoltaic power prediction has shown strong advantages. Deep learning models can automatically learn and identify complex patterns in large datasets, especially suitable for solving the volatility and uncertainty problems of photovoltaic power generation. Prediction methods based on deep learning include deep learning models such as Convolutional Neural Networks and Long Short Term Memory, showing significant performance advantages. Among them, Bidirectional Long Short-Term Memory further improves the prediction accuracy of the model by capturing bidirectional time series features. And deep learning methods also significantly improve the prediction accuracy of short-term and medium-long term predictions by integrating multi-source information (such as meteorological data, cloud maps).

[0005] However, there are still certain limitations in the existing technology for photovoltaic power prediction. These studies have not thoroughly explored seasonal characteristics. Although the model performs well in photovoltaic power prediction, the lack of in-depth analysis and utilization of seasonal characteristics may limit its prediction accuracy in different seasons. Seasonal factors play a crucial role in photovoltaic power generation. If these characteristics are not fully captured and modeled, it may affect the overall prediction performance of the model. In addition, these studies have not established the connection between meteorological information and seasonal changes. Photovoltaic power generation is greatly affected by meteorological conditions and is closely related to seasonal changes. The lack of in-depth exploration of the relationship between meteorological information and seasonality may lead to inaccurate predictions under changing weather conditions. Summary of the Invention

[0006] The object of the present invention is to provide a new energy power interval prediction method for mining the association between seasonal change characteristics and meteorological characteristics, including the following steps:

[0007] 1) Obtain the original photovoltaic power data.

[0008] 2) Decompose the original photovoltaic power data using the moving average algorithm to obtain a trend component and a seasonal component.

[0009] 3) Obtain meteorological data and construct a trend component prediction model and a seasonal component prediction model.

[0010] 4) Input the meteorological data and the seasonal component into the seasonal component prediction model to obtain a seasonal prediction result.

[0011] 5) Input the trend component into the trend component prediction model to obtain a trend prediction result.

[0012] 6) Calculate the sum of the seasonal prediction result and the trend prediction result to obtain the final photovoltaic prediction result.

[0013] Furthermore, the calculation formula for decomposing the original photovoltaic power data is as follows:

[0014] F (o) = T (a) + S (a) (1)

[0015] In the formula, F (o) is the original photovoltaic power data. S (a) is the seasonal component.

[0016] Among them, the trend component T (a) is as follows:

[0017]

[0018] Where \(i\) is the index of the original photovoltaic power data point, \(W\). s is the number of windows, and \(d\) is the window index.

[0019] Furthermore, the meteorological data includes temperature, relative humidity, global horizontal irradiance, and diffuse irradiance.

[0020] Furthermore, in step 4), the steps of obtaining the seasonal prediction result include:

[0021] 4.1) Input the meteorological data and the seasonal component into the first convolutional layer to obtain the first convolutional output, as follows:

[0022]

[0023] Where is the data of the first convolutional output. \(d\) c is the dilation coefficient, represents the weight of the convolutional kernel at position \(k\). \(k\) is the position, \(p\) is the position index. \(t\) is the time. is the data matrix of the photovoltaic power seasonal component data and the meteorological data.

[0024] 4.2) Normalize the data of the first convolutional output and perform a non-linear mapping on the normalized data using the ReLU activation function to obtain the activated data, as follows:

[0025]

[0026] Where is the activated data, is the normalized data.

[0027] 4.3) Input the activated data into the second convolutional layer and normalize the output of the second convolutional layer to obtain the input data

[0028] 4.4) Construct a hierarchical attention model, including a temporal attention model and a feature attention model.

[0029] The temporal attention model is used to calculate the importance of different time points.

[0030] The feature attention model is used to evaluate the weights of each feature at different time points.

[0031] 4.5) Input the input data into the hierarchical attention model to generate the output features of the hierarchical attention

[0032] 4.6) Construct a bidirectional long short-term memory network model and use the output features of the hierarchical attention as the input to the bidirectional long short-term memory network model to obtain the bidirectional long short-term memory network output data O b (t).

[0033] 4.7) Use a gated recurrent unit to capture the dynamic features of the bidirectional long short-term memory network output data O b (t) and generate a mapped output data through a linear layer

[0034] 4.8) Input the mapped output data into the hierarchical attention model to obtain the seasonal prediction result.

[0035] Furthermore, the time attention model is as follows:

[0036]

[0037] In the formula, is the input data. is the time-level weight, represents the time-level bias term. are the time-level attention network parameters. R a and R b are both learnable parameters of the time-level attention network. s is the time point index and t is the time. represents the calculation of the time attention score. represents the linear transformation and non-linear activation. represents the normalization of the attention score. is the output data of the time attention model.

[0038] Among them, the tanh function F (tanh) is as follows:

[0039] F (tanh) {x} = {exp(x) - exp(-x)} / {exp(x) + exp(-x)} (9)

[0040] In the formula, x is the input data of the tanh function.

[0041] Furthermore, the feature attention model is as follows:

[0042]

[0043] In the formula, is the input data. is the attention weight parameter. is the feature-level attention bias term. Represents the parameters of the feature-level attention network. F (tanh) Is the tanh function. R c And R d Are both learnable parameters of the feature-level attention network. s is the time point index and t is the time. Represents calculating the time attention score. Represents linear transformation and non-linear activation. Represents normalizing the attention score.

[0044] Furthermore, the output features of the hierarchical attention Are as follows:

[0045]

[0046] In the formula, s is the time point index and t is the time. Is the feature attention weight. Is the output data of the time attention model. F (HierarchicalAttention) Is the encapsulation formula of the hierarchical attention model. Is the input data.

[0047] Furthermore, the bidirectional long short-term memory network model includes two long short-term memory layers in opposite directions.

[0048] Both long short-term memory layers include an input gate, a forget gate, an output gate, a candidate memory unit, and a current memory unit, as follows:

[0049]

[0050] h b (t) = O b (t) * tanh(C b (t)) (19)

[0051] In the formula, I b (t) represents the output of the input gate at time t. And Are both learnable weights. Represents the hidden state at time t-1. H f (t) is the input sequence at time t. And Both represent learnable biases. F b (t) is the output of the forget gate at time t. O b (t) is the output of the output gate at time t. C b (t), C b (t-1) represent the unit state updates at times t and t-1 respectively. Represents the candidate unit state at time t-1. h b(t) represents the output of the hidden state at time t.

[0052] Among them, the activation function sigmoid is as follows:

[0053] sigmoid(y) = 1 / [1 + exp(-y)] (20)

[0054] In the formula, y is the input of the activation function sigmoid.

[0055] Furthermore, in step 5), the steps to obtain the trend prediction result are as follows:

[0056] 5.1) Construct a multi-layer perceptron model.

[0057] The multi-layer perceptron model includes an input layer, a hidden layer, and an output layer.

[0058] The input layer is used to receive the trend component.

[0059] The hidden layer is used to perform a non-linear transformation on the trend component.

[0060] The output result of the hidden layer is as follows:

[0061]

[0062] In the formula, is the output result of the j-th neuron in the hidden layer at time t. is the weight from the input layer to the hidden layer, i 1 is the data point index of the input layer, D is the number of data points in the input layer, T (a) [t] is the input trend component, is the bias term of the hidden layer, and σ is the activation function.

[0063] The output layer is used to output the trend prediction result.

[0064] 5.2) Use a multi-objective optimization algorithm to optimize the parameters of the multi-layer perceptron model.

[0065] 5.3) Use the multi-layer perceptron model with optimized parameters to perform trend prediction on the trend component to obtain the trend prediction result.

[0066] Furthermore, in step 5.2), the steps to optimize the parameters of the multi-layer perceptron model include:

[0067] 5.2.1) Randomly generate a population and set the iteration number m = 1.

[0068] Each individual in the population includes an initial learning rate, a weight decay coefficient, and the number of neurons.

[0069] 5.2.2) Divide the individuals in the population into two categories: discoverers and followers. The discoverers search for new individuals globally, and the followers search for new individuals locally around the discoverers and incorporate the new individuals into the population.

[0070] 5.2.3) Apply the m-th individual in the population to the multi-layer perceptron model for training and calculate the fitness value of the m-th individual.

[0071] The fitness value includes the mean square error.

[0072] 5.2.4) Update the positions of the individuals in the population according to the fitness value of the m-th individual, so that the individuals with lower fitness move towards the individuals with higher fitness.

[0073] 5.2.5) Determine whether the termination condition is reached. If so, output the optimal individual. If not, let m = m + 1 and return to step 5.2.3).

[0074] The termination conditions include reaching the maximum number of iterations and the convergence of the fitness value.

[0075] The technical effect of the present invention is beyond doubt. The present invention provides a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. This method allows meteorological information to affect the seasonal component of photovoltaic power generation data without affecting the trend component.

[0076] The present invention proposes a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The method includes a photovoltaic prediction strategy called "simplify seasonality, prioritize meteorology". This strategy aims to link meteorological information with the seasonal component of photovoltaic power generation data while preventing meteorology from affecting the trend component, thereby effectively reducing the impact of short-term seasonal meteorological fluctuations on the trend component of photovoltaic data. In addition, this study also proposes a seasonal component prediction unit with a double-layer hierarchical attention mechanism, which strengthens the attention to the connection between meteorological characteristics, key time nodes and seasonal components. These innovations enable the proposed prediction method to achieve higher prediction accuracy.

[0077] Using the photovoltaic data from 2018 to 2019 in Australia for testing, in the point prediction experiments with prediction lengths of 1 day, 2 days and 4 days, the prediction accuracy of the proposed prediction method exceeds 14 comparison models. In the interval prediction experiment, the accuracy of this model exceeds 18 comparison models. Description of the Drawings

[0078] Figure 1 is a schematic structural diagram of a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics of the present invention;

[0079] Figure 2It is the structure diagram of the internal seasonal term prediction unit of a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics according to the present invention;

[0080] Figure 3 It is the structure diagram of the internal trend term prediction unit of a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics according to the present invention;

[0081] Figure 4 It is the schematic diagram of the comparison of point prediction curves of a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics according to the present invention;

[0082] Figure 5 It is the schematic diagram of the interval prediction curve of a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics according to the present invention. Specific embodiments

[0083] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above technical idea of the present invention, various substitutions and changes made according to ordinary technical knowledge and customary means in the art shall be included within the protection scope of the present invention.

[0084] Embodiment 1:

[0085] See Figures 1 to 5 , a new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics, comprising the following steps:

[0086] 1) Obtain the original photovoltaic power data.

[0087] 2) Decompose the original photovoltaic power data by using the moving average algorithm to obtain the trend component and the seasonal component.

[0088] 3) Obtain the meteorological data and construct a trend component prediction model and a seasonal component prediction model.

[0089] 4) Input the meteorological data and the seasonal component into the seasonal component prediction model to obtain the seasonal prediction result.

[0090] 5) Input the trend component into the trend component prediction model to obtain the trend prediction result.

[0091] 6) Calculate the sum of the seasonal prediction result and the trend prediction result to obtain the final photovoltaic prediction result.

[0092] Embodiment 2:

[0093] A new - energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content is shown in Embodiment 1. Further, the calculation formula for decomposing the original photovoltaic power data is as follows:

[0094] F (o) =T (a) +S (a) (1)

[0095] In the formula, F (o) is the original photovoltaic power data. S (a) is the seasonal component.

[0096] Among them, the trend component T (a) is as follows:

[0097]

[0098] In the formula, i is the index of the original photovoltaic power data point, W s is the number of windows, and d is the window index.

[0099] Embodiment 3:

[0100] A new - energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content is shown in any one of Embodiments 1 to 2. Further, the meteorological data includes temperature, relative humidity, horizontal total radiation, and diffuse radiation.

[0101] Embodiment 4:

[0102] A new - energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content is shown in any one of Embodiments 1 to 3. Further, in step 4), the steps of obtaining the seasonal prediction result include:

[0103] 4.1) Input the meteorological data and the seasonal component into the first - layer convolutional layer to obtain the first - layer convolutional output, as follows:

[0104]

[0105] In the formula, is the data of the first - layer convolutional output. d c is the dilation coefficient, represents the weight of the convolutional kernel at position k. k is the position, p is the position index. t is the time. is the data matrix of the photovoltaic power seasonal component data and the meteorological data.

[0106] 4.2) For the data of the first - layer convolutional output Perform normalization processing, and use the ReLU activation function to perform non-linear mapping on the normalized data to obtain the activated data, as shown below:

[0107]

[0108] In the formula, is the activated data, is the data after normalization processing.

[0109] 4.3) Input the activated data into the second convolutional layer, and perform normalization processing on the output of the second convolutional layer to obtain the input data

[0110] 4.4) Construct a hierarchical attention model, including a temporal attention model and a feature attention model.

[0111] The temporal attention model is used to calculate the importance of different time points.

[0112] The feature attention model is used to evaluate the weights of each feature at different time points.

[0113] 4.5) Input the input data into the hierarchical attention model to generate the output features of hierarchical attention

[0114] 4.6) Construct a bidirectional long short-term memory network model, and input the output features of hierarchical attention into the bidirectional long short-term memory network model to obtain the bidirectional long short-term memory network output data O b (t).

[0115] 4.7) Use a gated recurrent unit to capture the dynamic features of the bidirectional long short-term memory network output data O b (t), and generate a mapped output data through a linear layer

[0116] 4.8) Input the mapped output data into the hierarchical attention model to obtain the seasonal prediction result.

[0117] Example 5:

[0118] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content can be found in any one of Examples 1 to 4. Further, the temporal attention model is as follows:

[0119]

[0120] In the formula, is the input data. is the time-level weight, represents the time-level bias term. is the parameter of the time-level attention network. R a and R b are both learnable parameters of the time-level attention network. s is the time point index and t is the time. represents calculating the time attention score. represents linear transformation and non-linear activation. represents normalizing the attention score. is the output data of the time attention model.

[0121] Among them, the tanh function F (tanh) is as follows:

[0122] F (tanh) {x} = {exp(x) - exp(-x)} / {exp(x) + exp(-x)} (9)

[0123] In the formula, x is the input data of the tanh function.

[0124] Example 6:

[0125] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content is shown in any one of Examples 1 to 5. Further, the feature attention model is as follows:

[0126]

[0127] In the formula, is the input data. is the attention weight parameter. is the feature-level attention bias term. represents the parameter of the feature-level attention network. F (tanh) is the tanh function. R c and R d are both learnable parameters of the feature-level attention network. s is the time point index and t is the time. represents calculating the time attention score. represents linear transformation and non-linear activation. represents normalizing the attention score.

[0128] Example 7:

[0129] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content is shown in any one of Examples 1 to 6. Further, the output features of the hierarchical attention are as follows:

[0130]

[0131] In the formula, s is the time point index and t is the time. is the feature attention weight. is the output data of the time attention model. F (HierarchoicalAttentoion) is the encapsulation formula of the hierarchical attention model. is the input data.

[0132] Embodiment 8:

[0133] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content can be found in any one of Embodiments 1 to 7. Further, the bidirectional long short-term memory network model includes two long short-term memory layers in opposite directions.

[0134] Both long short-term memory layers include an input gate, a forget gate, an output gate, a candidate memory unit, and a current memory unit, as follows:

[0135]

[0136] h b (t) = O b (t) * tanh(C b (t)) (19)

[0137] In the formula, I b (t) represents the output of the input gate at time t. and are both learnable weights. represents the hidden state at time t - 1. H f (t) is the input sequence at time t. and both represent learnable biases. F b (t) is the output of the forget gate at time t. O b (t) is the output of the output gate at time t. C b (t), C b (t - 1) represent the cell state updates at times t and t - 1 respectively. represents the candidate cell state at time t - 1. h b (t) represents the output of the hidden state at time t.

[0138] Among them, the activation function sigmoid is as follows:

[0139] sigmoid(y) = 1 / [1 + exp(-y)] (20)

[0140] In the formula, y is the input of the activation function sigmoid.

[0141] Example 9:

[0142] A new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content is as described in any one of Examples 1 to 8. Further, in step 5), the steps to obtain the trend prediction result are as follows:

[0143] 5.1) Construct a multi-layer perceptron model.

[0144] The multi-layer perceptron model includes an input layer, a hidden layer, and an output layer.

[0145] The input layer is used to receive the trend component.

[0146] The hidden layer is used to perform a non-linear transformation on the trend component.

[0147] The output result of the hidden layer is shown as follows:

[0148]

[0149] In the formula, is the output result of the j-th neuron in the hidden layer at time t. is the weight from the input layer to the hidden layer, i 1 is the data point index of the input layer, D is the number of data points in the input layer, T (a) [t] is the input trend component, is the bias term of the hidden layer, and σ is the activation function.

[0150] The output layer is used to output the trend prediction result.

[0151] 5.2) Optimize the parameters of the multi-layer perceptron model using a multi-objective optimization algorithm.

[0152] 5.3) Use the multi-layer perceptron model with optimized parameters to perform trend prediction on the trend component to obtain the trend prediction result.

[0153] Example 10:

[0154] A new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content is as described in any one of Examples 1 to 9. Further, in step 5.2), the steps to optimize the parameters of the multi-layer perceptron model include:

[0155] 5.2.1) Randomly generate a population and set the iteration number m = 1.

[0156] Each individual in the population includes an initial learning rate, a weight decay coefficient, and the number of neurons.

[0157] 5.2.2) Divide the individuals in the population into two categories: discoverers and followers. The discoverers search for new individuals globally, and the followers search for new individuals locally around the discoverers and incorporate the new individuals into the population.

[0158] 5.2.3) Apply the m-th individual in the population to the multi-layer perceptron model for training and calculate the fitness value of the m-th individual.

[0159] The fitness value includes the mean square error.

[0160] 5.2.4) Update the positions of the individuals in the population according to the fitness value of the m-th individual, so that the individuals with lower fitness move towards the individuals with higher fitness.

[0161] 5.2.5) Determine whether the termination condition is reached. If so, output the optimal individual. If not, let m = m + 1 and return to step 5.2.3).

[0162] The termination conditions include reaching the maximum number of iterations and the convergence of the fitness value.

[0163] Example 11:

[0164] See Figures 1 to 5 , a new energy power interval prediction method for mining the association between seasonal change characteristics and meteorological characteristics, comprising the following steps:

[0165] 1) Obtain the original photovoltaic power data.

[0166] 2) Decompose the original photovoltaic power data using the moving average algorithm to obtain the trend component and the seasonal component.

[0167] The moving average algorithm may face computational difficulties when the window does not fully contain enough data points. To solve this problem, the present invention adopts an endpoint processing method at the boundary position: when the window data is insufficient, a smaller window is automatically used for calculation to ensure the integrity of the trend component and the seasonal component.

[0168] Decompose the photovoltaic data into the trend component and the seasonal component, while the meteorological data (temperature, relative humidity, horizontal total radiation, diffuse radiation) is not decomposed. Through this decomposition strategy, different component characteristics of the photovoltaic power data can be processed to achieve higher prediction accuracy.

[0169] 3) Obtain the meteorological data and construct a trend component prediction model and a seasonal component prediction model.

[0170] In this embodiment, a strategy of "amplifying seasonality and giving priority to meteorology" is designed, which allows meteorological information to affect the seasonal component of photovoltaic power generation data while keeping the trend component unaffected by meteorological factors. In this invention, we combine the seasonal component of photovoltaic power data with meteorological data for the purpose of prediction. At the same time, the trend component of photovoltaic power data is analyzed independently without being integrated with meteorological data. This method can ensure accurate capture of seasonal changes affected by weather conditions, while the basic trend is not affected by short-term meteorological fluctuations.

[0171] 4) Input the meteorological data and the seasonal component into the seasonal component prediction model to obtain the seasonal prediction result.

[0172] 5) Input the trend component into the trend component prediction model to obtain the trend prediction result.

[0173] 6) Calculate the sum of the seasonal prediction result and the trend prediction result to obtain the final photovoltaic prediction result.

[0174] Embodiment 12:

[0175] A new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content can be seen in any one of Embodiments 11. Further, the calculation formula for decomposing the original photovoltaic power data is as follows:

[0176] F (o) = T (a) + S (a) (1)

[0177] In the formula, F (o) is the original photovoltaic power data. S (a) is the seasonal component.

[0178] Among them, the trend component T (a) is as follows:

[0179]

[0180] In the formula, i is the index of the original photovoltaic power data point, W s is the number of windows, and d is the window index.

[0181] Embodiment 13:

[0182] A new energy power interval prediction method for mining the correlation between seasonal change characteristics and meteorological characteristics. The main technical content can be seen in any one of Embodiments 11 to 12. Further, the meteorological data includes temperature, relative humidity, horizontal total radiation, and diffuse radiation.

[0183] Embodiment 14:

[0184] A new energy power interval prediction method for mining the association between seasonal change characteristics and meteorological characteristics. The main technical content can be found in any one of Embodiments 11 to 13. Further, in step 4), the steps to obtain the seasonal prediction result include:

[0185] 4.1) Input the meteorological data and seasonal components into the first convolutional layer to obtain the first convolutional output, as follows:

[0186]

[0187] In the formula, is the data of the first convolutional output. d c is the dilation coefficient, represents the weight of the convolutional kernel at position k. k is the position, p is the position index. t is the time. is the data matrix of the photovoltaic power seasonal component data and meteorological data.

[0188] 4.2) Normalize the data of the first convolutional output, and perform a non-linear mapping on the normalized data using the ReLU activation function to obtain the activated data, as follows:

[0189]

[0190] In the formula, is the activated data, is the data after normalization.

[0191] 4.3) Input the activated data into the second convolutional layer, and normalize the output of the second convolutional layer to obtain the input data

[0192] 4.4) Construct a hierarchical attention model, including a time attention model and a feature attention model.

[0193] The time attention model is used to calculate the importance of different time points.

[0194] The feature attention model is used to evaluate the weights of each feature at different time points.

[0195] 4.5) Input the input data into the hierarchical attention model to generate the output features of hierarchical attention

[0196] 4.6) Construct a bidirectional long short-term memory network model, and input the output features of hierarchical attention into the bidirectional long short-term memory network model to obtain the bidirectional long short-term memory network output data O b (t).

[0197] 4.7) Use a gated recurrent unit to capture the dynamic features of the output data O of the bidirectional long short-term memory network, and generate a mapped output data through a linear layer b (t), and generate a mapped output data through a linear layer

[0198] 4.8) Input the mapped output data into the hierarchical attention model to obtain the seasonal prediction result.

[0199] Example 15:

[0200] A new energy power interval prediction method for mining the correlation between seasonal change features and meteorological features. The main technical content can be found in any one of Examples 11 to 14. Further, the time attention model is as follows:

[0201]

[0202] In the formula, is the input data. is the time-level weight, represents the time-level bias term. are the time-level attention network parameters. R a and R b are both learnable parameters of the time-level attention network. s is the time point index, and t is the time. represents the calculation of the time attention score. represents the linear transformation and non-linear activation. represents the normalized attention score. is the output data of the time attention model.

[0203] Among them, the tanh function F (tanh) is as follows:

[0204] F (tanh) {x} = {exp(x) - exp(-x)} / {exp(x) + exp(-x)} (9)

[0205] In the formula, x is the input data of the tanh function.

[0206] Example 16:

[0207] A new energy power interval prediction method for mining the correlation between seasonal change features and meteorological features. The main technical content can be found in any one of Examples 11 to 15. Further, the feature attention model is as follows:

[0208]

[0209] In the formula, is the input data. is a learnable attention weight parameter. is a feature-level attention bias term. represents the feature-level attention network parameter. F (tanh) is the tanh function. R c and R d are both learnable parameters of the feature-level attention network. s is the time point index and t is the time. represents calculating the time attention score. represents linear transformation and non-linear activation. represents normalizing the attention score and obtaining the final representation output.

[0210] Embodiment 17:

[0211] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content can be found in any one of Embodiments 11 to 16. Further, the output features of the hierarchical attention are as follows:

[0212]

[0213] In the formula, s is the time point index and t is the time. is the feature attention weight. is the output data of the time attention model. F (HierarchicalAttention) is the encapsulation formula of the hierarchical attention model. is the input data.

[0214] Embodiment 18:

[0215] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content can be found in any one of Embodiments 11 to 17. Further, the bidirectional long short-term memory network model includes two long short-term memory layers in opposite directions.

[0216] Both long short-term memory layers include an input gate, a forget gate, an output gate, a candidate memory unit, and a current memory unit, as follows:

[0217]

[0218]

[0219] h b (t) = O b (t) * tanh(C b (t)) (19)

[0220] In the formula, I b (t) represents the output of the input gate at time t. and All are learnable weights. Represents the hidden state at time t-1. H f (t) is the input sequence at time t. and Both represent learnable biases. F b (t) is the output of the forget gate at time t. O b (t) is the output of the output gate at time t. C b (t), C b (t-1) represent the cell state updates at times t and t-1 respectively. Represents the candidate cell state at time t-1. h b (t) represents the output of the hidden state at time t.

[0221] Among them, the activation function sigmoid is as follows:

[0222] sigmoid(y) = 1 / [1 + exp(-y)] (20)

[0223] In the formula, y is the input of the activation function sigmoid.

[0224] Example 19:

[0225] A new energy power interval prediction method for mining the association between seasonal change features and meteorological features. The main technical content can be seen in any one of Examples 11 to 18. Further, the structural diagram of the internal trend term prediction unit of the prediction method is as Figure 3 shown. In step 5), the steps to obtain the trend prediction result are as follows:

[0226] 5.1) Construct a multi-layer perceptron model.

[0227] The multi-layer perceptron model includes an input layer, a hidden layer, and an output layer.

[0228] The input layer is used to receive the trend component.

[0229] The hidden layer is used to perform a non-linear transformation on the trend component.

[0230] The output result of the hidden layer is as follows:

[0231]

[0232] In the formula, is the output result of the j-th neuron in the hidden layer at time t. is the weight from the input layer to the hidden layer, i 1 is the data point index of the input layer, D is the number of data points in the input layer, T (a) [t] is the input trend component, is the bias term of the hidden layer, and σ is the activation function.

[0233] The output layer is used to output the trend prediction result.

[0234] 5.2) Use a multi-objective optimization algorithm to optimize the parameters of the multi-layer perceptron model.

[0235] 5.3) Use the multi-layer perceptron model with optimized parameters to perform trend prediction on the trend component to obtain the trend prediction result.

[0236] Embodiment 20:

[0237] A new energy power interval prediction method for mining the association between seasonal change characteristics and meteorological characteristics. The main technical content can be found in any one of Embodiments 11 to 19. Further, in step 5.2), the steps of optimizing the parameters of the multi-layer perceptron model include:

[0238] 5.2.1) Randomly generate a population and set the iteration number m = 1.

[0239] Each individual in the population includes an initial learning rate, a weight decay coefficient, and the number of neurons.

[0240] 5.2.2) Divide the individuals in the population into two categories: discoverers and followers. The discoverers search for new individuals globally, and the followers search for new individuals locally around the discoverers and incorporate the new individuals into the population.

[0241] 5.2.3) Apply the m-th individual in the population to the multi-layer perceptron model for training and calculate the fitness value of the m-th individual.

[0242] The fitness value includes the mean square error.

[0243] 5.2.4) Update the positions of the individuals in the population according to the fitness value of the m-th individual, so that the individuals with lower fitness move towards the individuals with higher fitness.

[0244] 5.2.5) Determine whether the termination condition is reached. If so, output the optimal individual. If not, set m = m + 1 and return to step 5.2.3).

[0245] The termination conditions include reaching the maximum number of iterations and the convergence of the fitness value.

[0246] The judgment condition for the convergence of fitness values is usually that when the change range of the best fitness value for several consecutive generations (e.g., 5 generations) is lower than a preset small threshold (such as 0.001), or the standard deviation of the fitness values of all individuals in the population is lower than a certain threshold (such as 0.005), it is considered that the fitness values have converged. For example, if in 5 consecutive generations, the best fitness value gradually changes from 0.2500 to 0.2510, and the change in each generation is less than 0.001, then the convergence condition is met, and the optimization process can be terminated and the current optimal individual can be output. This judgment method ensures that the optimization process stops in a timely manner when reaching a stable state, avoiding unnecessary waste of computing resources.

[0247] Example 21:

[0248] See Figures 1 to 5 , a new energy power interval prediction method for mining the association between seasonal change characteristics and meteorological characteristics, the main technical contents include:

[0249] This method allows meteorological information to affect the seasonal component of photovoltaic power generation data without affecting the trend component. The structural process of the photovoltaic power interval prediction method proposed in this example is as Figure 1 shown. The entire model process is as follows: the first section is decomposition, the second section is to amplify seasonality, giving priority to meteorological strategies, the third section is the seasonal term component prediction unit, the fourth section is the trend term component prediction unit, and the fifth section is the photovoltaic prediction output.

[0250] The first section: Decomposition

[0251] The present invention uses a moving average algorithm to decompose photovoltaic power data, decomposing the photovoltaic data into a trend component and a seasonal component, while the meteorological data (temperature, relative humidity, horizontal total radiation, diffuse radiation) is not decomposed. Through this decomposition strategy, different component characteristics of photovoltaic power data can be processed, thereby achieving higher prediction accuracy.

[0252] Establish the following decomposition formula:

[0253] F (o) = T (a) + S (a)

[0254] Among them, F (o) represents the original photovoltaic power data, in kilowatts (kW), and is sourced from the photovoltaic power dataset. T (a) represents the trend component, obtained by calculating the original photovoltaic power data, with the unit of kilowatts (kW). S (a) represents the seasonal component, obtained by calculating the original photovoltaic power data, with the unit of kilowatts (kW).

[0255] The formula for obtaining the trend component through the moving average algorithm is as follows:

[0256]

[0257] where i is the index of the current data point, and W s is the number of windows, which is set to 10 in the present invention.

[0258] The moving average algorithm may face computational difficulties when the window does not fully contain enough data points. To solve this problem, the present invention adopts an endpoint processing method at the boundary position: when the window data is insufficient, a smaller window is automatically used for calculation to ensure the integrity of the trend component and the seasonal component.

[0259] After obtaining the trend component T (a) , the seasonal component can be calculated through the following formula:

[0260] S (a) = F (o) - T (a)

[0261] Section 2: Amplifying seasonality, giving priority to meteorological strategies

[0262] A strategy of "amplifying seasonality, giving priority to meteorology" is designed, allowing meteorological information to affect the seasonal component of photovoltaic power generation data while keeping the trend component unaffected by meteorological factors. In this invention, we combine the seasonal component of photovoltaic power data with meteorological data for prediction purposes. At the same time, the trend component of photovoltaic power data is analyzed independently without integration with meteorological data. This method can ensure accurate capture of seasonal changes affected by weather conditions, while the basic trend is not affected by short-term meteorological fluctuations.

[0263] Section 3: Seasonal term component prediction unit

[0264] The input data of the seasonal component prediction unit includes two parts: the seasonal component of photovoltaic power data and meteorological data (including temperature, relative humidity, global horizontal irradiance, diffuse irradiance). This unit aims to explore the dependence relationship between photovoltaic power and meteorological data. A double-layer temporal convolutional structure is mainly designed, and an instance normalization strategy is introduced after the convolutional layer. At the same time, a hierarchical attention mechanism, including meteorological feature attention and time-specific attention, is embedded to improve the prediction accuracy. The internal structure diagram of the seasonal term prediction unit of the prediction method is as Figure 2 shown.

[0265] The input data is a composite data matrix [S (a) , M (f) , where: S (a) is the seasonal component of photovoltaic power data; M (f)is meteorological data, including temperature, relative humidity, total horizontal radiation, and diffuse radiation. Temperature is in degrees Celsius (°C), relative humidity is in percentage (%), and radiation is in watts per square meter (W / m 2 ). The meteorological data is sourced from sensor collection and is represented as pure numerical values.

[0266] The input data [S (a) ,M (f) first enters the first layer of convolution in the temporal convolutional unit, and the calculation formula is as follows:

[0267]

[0268] where represents the weight of the convolutional kernel at position k and is represented as a pure numerical value. is the output of the first layer of causal convolution and is represented as a pure numerical value. d c is the dilation coefficient, represented as a pure numerical value, used to enable the model to capture dependencies within a larger range.

[0269] The data after convolution undergoes instance normalization to accelerate training and improve the generalization ability of the model, resulting in the data after instance normalization Subsequently, the ReLU activation function is used for non-linear mapping, and the calculation formula is as follows:

[0270]

[0271] The data after ReLU activation enters the second layer of convolution and instance normalization, finally forming the input of the hierarchical attention mechanism

[0272] The hierarchical attention mechanism consists of temporal attention and feature attention: Temporal attention calculates the importance of different time points; Feature attention evaluates the weights of each feature at different time points.

[0273] The formula for temporal attention is as follows:

[0274]

[0275] where is represented as a pure numerical value and represents the input data. is also represented as a pure numerical value and represents the temporal-level weight. represents the temporal-level bias term and is represented as a pure numerical value. The represented as a pure numerical value represents the parameter set to 1 in the attention network. R a and R b are both represented as pure numerical values and are the learnable parameters of the temporal-level attention network. The is the obtained attention weight, represented by pure numerical values is the output data after weighting. F represented by pure numerical values (tanh) Describes the tanh function, and the equation of the tanh function is as follows.

[0276] F (tanh) {x} = {exp(x) - exp(-x)} / {exp(x) + exp(-x)}

[0277] The feature attention formula is as follows:

[0278]

[0279] where is represented by pure numerical values and describes the input data. is also represented by pure numerical values and represents the feature-level attention weight. Represented by pure numerical values is the feature-level attention bias term. Indicates that the parameter setting in the attention network is 1. R c and R d are both represented by pure numerical values and are the learnable parameters of the feature-level attention network. Represented by pure numerical values Indicates the obtained attention weight.

[0280] Finally, through weighted calculation, the output feature formula for hierarchical attention is as follows:

[0281]

[0282] where represented by pure numerical values represents the final weighted feature representation, which contains the importance of the time dimension and the feature dimension. Considering the subsequent use of the hierarchical attention mechanism, the following formula is used to encapsulate the formula:

[0283]

[0284] The processed data is input into a bidirectional long short-term memory network (BiLSTM). This network consists of two long short-term memory layers in opposite directions, running along the time series from the past to the present and from the present to the past respectively. Each unidirectional long short-term memory layer includes an input gate, a forget gate, an output gate, a candidate memory unit, and a current memory unit, and the formula is as follows:

[0285]

[0286] h b (t) = O b (t) * tanh(C b (t))

[0287] where Ib (t) is represented by a pure numerical value, indicating the output of the input gate at the current time step t. F b (t) is also represented by a pure numerical value, indicating the output of the forget gate. O b (t) is represented by a pure numerical value, indicating the output of the output gate. C represented by a pure numerical value b (t) shows the storage unit at the current time step \(t\), while C represented by a pure numerical value shows the storage unit at the previous time step (t - 1). h b (t) is represented by a pure numerical value, representing the output of the hidden state at time step h b (t). and are both represented by pure numerical values, indicating learnable weights. Similarly, represented by pure numerical values and indicate learnable biases. H represented by a pure numerical value f (t) depicts the input sequence at time step t, while H represented by a pure numerical value is a condensation of the hidden state at the previous time step (t - 1).

[0288] "tanh" and "sigmoid" represent activation functions, with the following formulas:

[0289] tanh(x) = {exp(x) - exp(-x)} / {exp(x) + exp(-x)}

[0290] sigmoid(x) = 1 / [1 + exp(-x)]

[0291] The output data O b (t) uses a gated recurrent unit to capture the dynamic features of the sequence and maps these features to a new feature space through a linear layer. The linear mapping is represented by a pure numerical value, and the gated recurrent unit uses the same parameters as the bidirectional long short-term memory, including learnable weights and biases represented by pure numerical values, and produces a new mapped output also represented by a pure numerical value.

[0292] The mapped data is input into a hierarchical attention mechanism, and finally the predicted output of the seasonal term prediction unit is obtained

[0293] Section 4: Trend Term Component Prediction Unit

[0294] This method first extracts the trend component of photovoltaic power data, and then uses a Multilayer Perceptron model for trend prediction and a multi-objective optimization algorithm to optimize the model parameters. The structure diagram of the internal trend term prediction unit of the prediction method is as shown in Figure 3 shown.

[0295] The Multilayer Perceptron model consists of an input layer, a hidden layer, and an output layer. The input layer receives the trend component data, the hidden layer performs a non-linear transformation on the data, and the output layer gives the final trend prediction result. The output of the hidden layer is calculated by the following formula:

[0296]

[0297] where is the output of the j-th neuron in the hidden layer, is the weight from the input layer to the hidden layer, DDD is the number of data points in the input layer, T (a) [t] is the input trend component data, is the bias term of the hidden layer, and σ is the activation function.

[0298] Since this study needs to find the optimal combination of multiple parameters, we extend the Sparrow Search Algorithm to a multi-objective Sparrow Search Algorithm and apply this algorithm to optimize three key parameters in the trend prediction unit: learning rate, weight decay coefficient, and number of neurons. The specific optimization process is as follows:

[0299] 1) Population initialization: First, randomly generate a population, where each individual represents a parameter combination, including the initial learning rate, weight decay coefficient, and number of neurons. 2) Grouping and fitness calculation: In the population, some individuals act as discoverers, responsible for searching for new parameter combinations globally, while other individuals act as followers, performing local searches around the discoverers. Each parameter combination is evaluated through a fitness function to determine its effectiveness in model training. 3) Update and iteration: In each iteration, the three key parameters (learning rate, weight decay coefficient, and number of neurons) are applied to the model for training, and their fitness values are calculated (in this study, the fitness function is set to mean squared error). The population position is updated according to the fitness values, making the individuals with lower fitness move towards the individuals with higher fitness, thus gradually approaching the optimal parameter combination. 4) Termination condition: When the maximum number of iterations is reached or the fitness converges, the algorithm outputs the optimal parameter combination, including the learning rate, weight decay coefficient, and number of neurons.

[0300] Then, the trend prediction unit is trained using the optimal parameter combination, and finally, a trend prediction output is generated The output is a pure numerical value.

[0301] Section 5: Photovoltaic Prediction Output

[0302] The prediction output of the seasonal prediction unit is represented by pure numerical values. The prediction output of the trend prediction unit is also represented by pure numerical values. The final photovoltaic prediction output is represented by pure numerical values and is calculated as and the sum of.

[0303] Using the photovoltaic data from 2018 to 2019 in Australia for testing, in the point prediction experiments with prediction lengths of 1 day, 2 days, and 4 days, the prediction accuracy of the proposed prediction method exceeded 14 comparison models. In the interval prediction experiment, the accuracy rate of this model exceeded 18 comparison models. The schematic diagram of the comparison of the first prediction curve of 384 steps of point prediction by the interval prediction method of the present invention is as Figure 4 shown. The schematic diagram of the interval prediction curve of the interval prediction method is as Figure 5 shown.

Claims

1. A new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics, characterized in that: The following steps are involved: 1) Obtaining raw photovoltaic power data; 2) The moving average algorithm is used to decompose the original photovoltaic power data to obtain the trend component and seasonal component. 3) Obtain meteorological data and construct trend component prediction models and seasonal component prediction models; 4) Input the meteorological data and seasonal components into the seasonal component prediction model to obtain seasonal prediction results; 5) Input the trend component into the trend component prediction model to obtain the trend prediction result. 6) Calculate the sum of seasonal forecast results and trend forecast results to obtain the final photovoltaic forecast results.

2. The new energy power interval prediction method for mining seasonal variation characteristics and meteorological characteristics according to claim 1 is characterized in that: The calculation formula for decomposing the original photovoltaic power data is as follows: F (o) =T (a) +S (a) (1) In the formula, F (o) is the original photovoltaic power data; S (a) is the seasonal component; Among them, the trend component T (a) As shown below: Where i is the index of the original photovoltaic power data point, W s is the number of windows, and d is the window index.

3. The new energy power interval prediction method according to claim 1, characterized in that: The meteorological data include temperature, relative humidity, horizontal global radiation, and diffuse radiation.

4. The new energy power interval prediction method for mining seasonal variation characteristics and meteorological characteristics according to claim 1 is characterized in that: In step 4), the steps of obtaining seasonal forecast results include: 4.1) Input the meteorological data and seasonal components into the first convolution layer to obtain the first convolution output as shown below: In the formula, The data output by the first layer of convolution; d c is the expansion coefficient, represents the weight of the convolution kernel at position k; k is the position, p is the position index; t is the time; It is the data matrix of photovoltaic power seasonal component data and meteorological data; 4.2) Output data of the first layer of convolution Normalize the data and use the ReLU activation function to perform nonlinear mapping on the normalized data to obtain activated data, as shown below: In the formula, For the activated data, is the data after normalization; 4.3) Input the activated data into the second convolutional layer, and normalize the output of the second convolutional layer to obtain the input data 4.4) Construct a hierarchical attention model, including a temporal attention model and a feature attention model; The temporal attention model is used to calculate the importance of different time points; The feature attention model is used to evaluate the weight of each feature at different time points; 4.5) Input data Input hierarchical attention model to generate output features of hierarchical attention 4.6) Build a bidirectional long short-term memory network model and use the output features of the hierarchical attention Input the bidirectional long short-term memory network model and obtain the bidirectional long short-term memory network output data O b (t); 4.7) Using gated recurrent units to capture bidirectional long short-term memory network output data b (t) and generates mapping output data through a linear layer 4.8) Map the output data Input the hierarchical attention model to obtain seasonal prediction results.

5. The new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics according to claim 4 is characterized in that: The temporal attention model is as follows: In the formula, For input data; is the time-level weight, represents the time-level deviation term; is the temporal attention network parameter; R a and R b are all learnable parameters of the time-level attention network; s is the time point index, t is the time; Indicates the calculation time attention score; Represents linear transformation and nonlinear activation; represents the normalized attention score; is the output data of the temporal attention model; Among them, the tanh function F (tanh) As shown below: F (tanh) {x}={exp(x)-exp(-x)} / {exp(x)+exp(-x)} (9) Where x is the input data of the tanh function.

6. The new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics according to claim 4 is characterized in that: The feature attention model is as follows: In the formula, For input data; is the attention weight parameter; is the feature-level attention bias term; represents the feature-level attention network parameters; F (tanh) is the tanh function; R c and R d are all learnable parameters of the feature-level attention network; s is the time point index, t is the time; Indicates the calculation time attention score; Represents linear transformation and nonlinear activation; represents the normalized attention score.

7. The new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics according to claim 4 is characterized in that: The output features of the level attention As shown below: In the formula, s is the time point index, and t is the time; is the feature attention weight; is the output data of the temporal attention model; F (HierarchicalAttention) It is the encapsulation formula of the hierarchical attention model; For input data.

8. The method for predicting new energy power intervals by mining seasonal variation characteristics and meteorological characteristics according to claim 4, characterized in that: The bidirectional long short-term memory network model includes two long short-term memory layers in opposite directions; Both long short-term memory layers include input gate, forget gate, output gate, candidate memory unit and current memory unit, as shown below: h b (t)=O b (t)*tanh(C b (t)) (19) In the formula, I b (t) represents the output of the input gate at time t; and All are learnable weights; represents the hidden state at time t-1; H f (t) is the input sequence at time t; and Both represent learnable bias; F b (t) is the output of the forget gate at time t; b (t) is the output of the output gate at time t; C b (t), C b (t-1) represents the unit state update at time t and t-1 respectively; represents the candidate unit state at time t-1; h b (t) represents the output of the hidden state at time t; Among them, the activation function sigmoid is as follows: sigmoid(y)=1 / [1+exp(-y)] (20) Where y is the input of the activation function sigmoid.

9. The new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics according to claim 1 is characterized in that: In step 5), the steps to obtain the trend prediction results are as follows: 5.1) Construct a multi-layer perceptron model; The multi-layer perceptron model includes an input layer, a hidden layer and an output layer; The input layer is used to receive the trend component; The hidden layer is used to perform nonlinear transformation on the trend component; The output of the hidden layer is as follows: In the formula, is the output result of the jth neuron in the hidden layer at time t; is the weight from the input layer to the hidden layer, i1 is the data point index of the input layer, D is the number of data points in the input layer, T (a) [t] is the trend component of the input, is the bias term of the hidden layer, σ is the activation function; The output layer is used to output trend prediction results; 5.2) Use a multi-objective optimization algorithm to optimize the parameters of the multi-layer perceptron model; 5.3) Use the multi-layer perceptron model with optimized parameters to predict the trend component and obtain the trend prediction results.

10. The new energy power interval prediction method for mining the association between seasonal variation characteristics and meteorological characteristics according to claim 9 is characterized in that: In step 5.2), the steps of optimizing the parameters of the multilayer perceptron model include: 5.2.1) Randomly generate a population and set the number of iterations m = 1; Each individual in the population includes the initial learning rate, weight decay coefficient, and number of neurons; 5.2.2) Divide the individuals in the group into two categories: discoverers and followers. Discoverers search for new individuals globally, and followers search for new individuals locally around the discoverers and incorporate the new individuals into the group. 5.2.3) Apply the mth individual in the group to the multilayer perceptron model for training and calculate the fitness value of the mth individual; The fitness value includes a mean square error; 5.2.4) According to the fitness value of the mth individual, update the position of the individuals in the group so that the individuals with lower fitness move towards the individuals with higher fitness; 5.2.5) Determine whether the termination condition is met. If so, output the optimal individual. If not, set m=m+1 and return to step 5.2.3); The termination conditions include reaching the maximum number of iterations and the fitness value converging.

Images