New energy power generation prediction method and system based on improved LSTM model

Abstract

The present invention discloses a method and system for predicting renewable energy power generation based on an improved LSTM model, relating to the technical field of renewable energy power generation prediction. The method comprises: obtaining historical power generation data and multidimensional meteorological grid data of a renewable energy power station, performing preprocessing and feature extraction to obtain preprocessed data; using the preprocessed data as input, using a trained improved LSTM model to predict the power generation of the renewable energy power station to obtain a power generation forecast value; wherein the improved LSTM model includes a periodic memory gating unit that introduces a periodic weight term; during the training process of the improved LSTM model, a memory decay compensation mechanism is introduced, and cross-validation and grid search methods are used to optimize the architecture and hyperparameters. The present invention alleviates the technical problems of insufficient prediction accuracy and poor model stability that exist in traditional prediction methods when dealing with long-term trends, seasonal changes, and complex meteorological influences.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation prediction, and in particular to a renewable energy power generation prediction method and system based on an improved LSTM model. Background Art

[0002] In recent years, with the large-scale integration of renewable energy sources such as wind power and photovoltaics, renewable energy power generation data has exhibited strong nonlinear, periodic, and time-varying characteristics. Traditional forecasting methods (including regression models, support vector machines, and conventional LSTM networks) generally suffer from insufficient forecast accuracy and poor model stability when dealing with long-term trends, seasonal variations, and complex meteorological influences. Conventional LSTMs are particularly poor at capturing long-term dependencies, periodicity, and multi-scale information, making it difficult to meet the high-precision forecasting requirements for renewable energy power system scheduling and planning. Summary of the Invention

[0003] The purpose of the present invention is to provide a new energy power generation prediction method and system based on an improved LSTM model in order to solve at least one of the above technical problems.

[0004] In the first aspect, an embodiment of the present invention provides a new energy power generation prediction method based on an improved LSTM model, which is applied to a new energy power station; the method comprises: obtaining historical power generation data and multidimensional meteorological grid data of the new energy power station; preprocessing and feature extraction of the historical power generation data and the multidimensional meteorological grid data to obtain preprocessed data; using the preprocessed data as input, and using a trained improved LSTM model to predict the power generation of the new energy power station to obtain a power generation prediction value; wherein, the improved LSTM model comprises an input gate, a forget gate and an output gate, and the input gate, the forget gate and the output gate are all periodic memory gated units that introduce a periodic weight term; during the training process of the improved LSTM model, a memory decay compensation mechanism is introduced and cross-validation and grid search methods are used for architecture and hyperparameter optimization.

[0005] Furthermore, the historical power generation data includes the historical power generation time series data of the new energy power station; the multidimensional meteorological grid data includes wind speed, temperature, humidity, and irradiance.

[0006] Furthermore, the historical power generation data and the multidimensional meteorological grid data are preprocessed and feature extracted to obtain preprocessed data, including: data cleaning and normalization processing of the historical power generation data and the multidimensional meteorological grid data to obtain normalized historical power generation data and normalized multidimensional meteorological grid data, respectively; local spatial feature extraction of the normalized multidimensional meteorological raster data based on a spatiotemporal convolutional network to obtain multidimensional meteorological spatial feature data; fusion of the multidimensional meteorological spatial feature data with the normalized historical power generation data to obtain fused data; and dimensionality reduction and screening of the fused data based on a principal component analysis method to obtain preprocessed data.

[0007] Furthermore, the structural formula of the input gate in the improved LSTM model includes:

[0008]

[0009] Where: Indicates Input the activation value of the gate at any moment; Represents the Sigmoid activation function; represents the weight matrix of the input gate; express The hidden state of the moment; express Input data at the moment; represents the bias of the input gate; Represents the periodic weight term introduced in the input gate; Represents a time variable or a moment index; the structural formula of the forget gate in the improved LSTM model includes:

[0010]

[0011] Where: Indicates The activation value of the momentary forget gate; Represents the weight matrix of the forget gate; Represents the bias of the forget gate; Represents the periodic weight term introduced in the forget gate; represents the memory decay compensation term; the structural formula of the output gate in the improved LSTM model includes:

[0012]

[0013] Where: Indicates Output the activation value of the gate at that moment; Represents the weight matrix of the output gate; Represents the bias of the output gate; Represents the periodic weight term introduced in the output gate.

[0014] Furthermore, the unit state update formula of the improved LSTM model includes:

[0015]

[0016] Where: Indicates The unit state at the moment; represents the cell state at time t-1; represents element-wise product; represents the hyperbolic tangent activation function; represents the weight matrix used to calculate the candidate unit state; Represents the corresponding bias; the hidden state update formula of the improved LSTM model includes:

[0017]

[0018] Where: express The hidden state of the moment.

[0019] Furthermore, the memory decay compensation term is adaptively adjusted according to the second-order derivative information of the loss function of the improved LSTM model with respect to the time step during the training process; the calculation formula of the memory decay compensation term includes:

[0020]

[0021] Where: Indicates The memory decay compensation term at the moment; is the adjustment coefficient; Represents the loss function used by the improved LSTM model during training.

[0022] Furthermore, the period weight term includes:

[0023]

[0024] in, Indicates The period weight item corresponding to the moment; is a learnable scaling parameter used to adjust the amplitude of the cycle weight during the learning process; Respectively Fourier cosine coefficients and sine coefficients corresponding to the periodic components; represents the fundamental frequency of the nth periodic component; is the index of the periodic component; represents the total number of periodic components considered.

[0025] In the second aspect, an embodiment of the present invention also provides a new energy power generation prediction system based on an improved LSTM model, which is applied to a new energy power station; it includes: an acquisition module, a preprocessing module and a prediction module; wherein the acquisition module is used to obtain the historical power generation data and multidimensional meteorological grid data of the new energy power station; the preprocessing module is used to preprocess and extract features of the historical power generation data and the multidimensional meteorological grid data to obtain the preprocessed data; the prediction module is used to use the preprocessed data as input, and use the trained improved LSTM model to predict the power generation of the new energy power station to obtain a power generation prediction value; wherein the improved LSTM model includes an input gate, a forget gate and an output gate, and the input gate, the forget gate and the output gate are all periodic memory gated units that introduce periodic weight terms; in the training process of the improved LSTM model, a memory decay compensation mechanism is introduced and cross-validation and grid search methods are used for architecture and hyperparameter optimization.

[0026] In a third aspect, an embodiment of the present invention further provides an electronic device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the method provided in the embodiment of the present invention when executing the computer program.

[0027] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method provided in the embodiment of the present invention is implemented.

[0028] This paper provides a method and system for forecasting renewable energy power generation based on an improved LSTM model. By introducing periodic weights and a memory decay compensation mechanism into the traditional LSTM model, this method fully captures the periodicity and long-term dependency information in historical power generation and meteorological data, thereby improving the accuracy of medium- and long-term forecasts of renewable energy power generation. Furthermore, by extracting spatial features and selecting key features from multidimensional meteorological data, this method effectively reduces redundant information, improves the model's training efficiency and generalization capabilities, and alleviates the technical issues of insufficient forecast accuracy and poor model stability that plague traditional forecasting methods when dealing with long-term trends, seasonal variations, and complex meteorological influences. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the implementation methods of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific implementation methods or the description of the prior art. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0030] Figure 1 A flowchart of a new energy power generation prediction method based on an improved LSTM model provided in an embodiment of the present invention;

[0031] Figure 2 A schematic diagram of a new energy power generation prediction system based on an improved LSTM model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] Example 1

[0034] Figure 1 This is a flowchart of a new energy power generation prediction method based on an improved LSTM model provided in accordance with an embodiment of the present invention, which is applied to new energy power stations. Figure 1 As shown, the method specifically includes the following steps:

[0035] Step S102: Acquire historical power generation data and multi-dimensional meteorological grid data of the new energy power station.

[0036] Specifically, the historical power generation data includes the historical power generation time series data of the new energy power station; for example, the monthly or annual power generation data of the new energy power station in the past 1 to 3 years, including the specific timestamp and the corresponding power generation value.

[0037] Multidimensional meteorological raster data includes wind speed, temperature, humidity, and irradiance. This data corresponds to historical power generation time series data. This data is provided in raster form, covers the spatial extent of the region where the new energy power station is located, and has the same temporal resolution as the historical power generation time series data (e.g., monthly or annual).

[0038] Optionally, the historical power generation time series data can be obtained from the power grid dispatching system or the monitoring system of the new energy power station; the meteorological data can be obtained from the meteorological data platform provided by the meteorological department or an additional meteorological data service company.

[0039] Step S104: preprocess and extract features from historical power generation data and multi-dimensional meteorological grid data to obtain preprocessed data. Specifically, the steps include:

[0040] Step S1041 , performing data cleaning and normalization processing on the historical power generation data and the multi-dimensional meteorological grid data, to obtain normalized historical power generation data and normalized multi-dimensional meteorological grid data respectively.

[0041] Specifically, data cleaning includes removing outliers and missing values ​​in the data; missing values ​​can be filled by interpolation methods (such as linear interpolation or time series interpolation); outliers can be eliminated or corrected by setting a reasonable threshold range.

[0042] Specifically, in order to ensure the consistency and integrity of the timestamps of the data, the embodiment of the present invention also performs alignment processing on the data so that the historical power generation data and the multi-dimensional meteorological grid data correspond one to one in time.

[0043] Specifically, the normalization processing formula includes:

[0044]

[0045] in, Represents the original data value; Represents the normalized data value; Represents a data set The minimum value of Represents a data set The data range after normalization is [0,1], which helps improve the training efficiency and stability of the model.

[0046] Step S1042: extract local spatial features from the normalized multidimensional meteorological raster data based on a spatiotemporal convolutional network to obtain multidimensional meteorological spatial feature data.

[0047] Specifically, the meteorological raster data is input into the spatiotemporal convolutional network (STCNN), and the data is subjected to feature extraction through the convolution layer and pooling layer to extract the feature map containing local spatial information.

[0048] Step S1043 , fusing the multi-dimensional meteorological spatial feature data with the normalized historical power generation data to obtain fused data.

[0049] Optionally, the fusion method may adopt a simple concatenation or weighted summation method, which depends on the characteristics of the data and the requirements of the model, and is not specifically limited in the present invention.

[0050] Step S1044: perform dimensionality reduction and screening on the fused data based on the principal component analysis method to obtain preprocessed data.

[0051] Specifically, principal component analysis (PCA) calculates eigenvalues ​​and eigenvectors and selects the top k principal components as the model's input features. The value of k is determined based on the cumulative contribution rate (for example, 95%). This dimensionality-reduced feature data retains the primary information of the original data while reducing redundant information, improving model training efficiency and predictive effectiveness.

[0052] In step S106 , the pre-processed data is used as input, and the trained improved LSTM model is used to predict the power generation of the new energy power station to obtain a power generation prediction value.

[0053] Among them, the improved LSTM model includes input gate, forget gate and output gate. The input gate, forget gate and output gate are all periodic memory gating units that introduce periodic weight terms. During the training process of the improved LSTM model, a memory decay compensation mechanism is introduced and cross-validation and grid search methods are used to optimize the architecture and hyperparameters.

[0054] Specifically, the periodic weight term is obtained by frequency decomposing the historical state sequence using Fourier basis functions, which is used to capture the annual, quarterly and monthly periodic characteristics, and at the same time, the periodic components are weighted and fused through learnable parameters.

[0055] The embodiment of the present invention introduces a memory decay compensation mechanism in the training process of the improved LSTM model, and dynamically adjusts the forget gate coefficient according to the second-order derivative information of the loss function to compensate for the decay of long-term dependent information when it is transmitted across cycles.

[0056] The embodiment of the present invention adopts cross-validation and grid search methods to optimize and improve the architecture of each layer and hyperparameters of the LSTM model, ensuring that the model skills accurately fit historical data and have good generalization ability.

[0057] Specifically, the improved LSTM model uses cross-validation and grid search methods to determine the weights, biases, periodic weight term parameters and memory decay compensation coefficients of each layer during the training process, thereby ensuring that the model has high accuracy and robustness in the medium and long-term prediction stage while fitting historical data.

[0058] Specifically, in an embodiment of the present invention, the structural formula of the input gate in the improved LSTM model includes:

[0059]

[0060] Where: Indicates Input the activation value of the gate at any moment; Represents the Sigmoid activation function; represents the weight matrix of the input gate; express The hidden state of the moment; express Input data at the moment; represents the bias of the input gate; Represents the periodic weight term introduced in the input gate; Represents a time variable or moment index;

[0061] The structural formula for improving the forget gate in the LSTM model includes:

[0062]

[0063] Where: Indicates The activation value of the momentary forget gate; Represents the weight matrix of the forget gate; Represents the bias of the forget gate; Represents the periodic weight term introduced in the forget gate, which is calculated by Fourier basis function and is constructed in the same way as the periodic weight term in the input gate; Represents the memory decay compensation term, which is adaptively adjusted according to the second-order derivative information of the loss function with respect to time during model training to compensate for the decay of long-term dependent information;

[0064] The structural formula for improving the output gate in the LSTM model includes:

[0065]

[0066] Where: Indicates Output the activation value of the gate at that moment; Represents the weight matrix of the output gate; Represents the bias of the output gate; Represents the periodic weight term introduced in the output gate.

[0067] Specifically, the period weight items include:

[0068]

[0069] in, Indicates The period weight item corresponding to the moment; is a learnable scaling parameter used to adjust the amplitude of the cycle weight during the learning process; Respectively Fourier cosine coefficients and sine coefficients corresponding to the periodic components; Represents the fundamental frequency of the nth periodic component (corresponding to annual, quarterly, monthly, etc. periods); is the index of the periodic component, ; represents the total number of periodic components considered.

[0070] For example, the cycle weight term of the input gate is:

[0071]

[0072] Where, Represents the periodic weight term introduced in the input gate; represents a learnable scaling parameter in the input gate that controls the amplitude of the periodic term.

[0073] The cycle weight term of the output gate is:

[0074]

[0075] Where, represents the periodic weight term introduced in the output gate; A learnable scaling parameter representing the amplitude of the output gate control periodic term; Respectively The periodic components correspond to the Fourier cosine and sine coefficients in the output gate.

[0076] For example, consider the year , quarter , monthly These three periodic components , through learnable parameters The components of each period are weighted and integrated to effectively capture the annual, quarterly and monthly periodic characteristics.

[0077] Specifically, the memory decay compensation term is adaptively adjusted based on the second-order derivative information of the loss function with respect to the time step during the training process of the improved LSTM model. The calculation formula of the memory decay compensation term includes:

[0078]

[0079] Where: Indicates The memory decay compensation term at the moment; is the adjustment coefficient, used to control the compensation amplitude; Represents the loss function used by the improved LSTM model during training; Represents the loss function About Time The second-order derivative of , which reflects the curvature of the loss over time, is used to dynamically optimize the adjustment of the forget gate.

[0080] Specifically, in the early stages of training, if the curvature of the loss function over time is large (i.e., the absolute value of the second-order derivative is large), it means that the model prediction error changes rapidly. In this case, it is appropriate to increase The value of , strengthens the compensation for long-term dependent information; as the training progresses, if the loss function tends to be stable and the absolute value of the second-order derivative becomes smaller, then the corresponding reduction Through this adaptive adjustment, the attenuation of long-term dependency information when it is transmitted across cycles is compensated.

[0081] Specifically, in an embodiment of the present invention, the unit state update formula of the improved LSTM model includes:

[0082]

[0083] Where: Indicates The unit state at the moment; represents the cell state at time t-1; represents the element-wise product, i.e. the Hadamard product; represents the hyperbolic tangent activation function; represents the weight matrix used to calculate the candidate unit state; Represents the corresponding bias; i t Used to control the ratio of new information added.

[0084] Specifically, the hidden state update formula of the improved LSTM model includes:

[0085]

[0086] Where: express The hidden state of the moment, Used to modulate the result of nonlinear transformation of the unit state.

[0087] The method provided by an embodiment of the present invention further includes training the improved LSTM model to obtain a trained improved LSTM model. Specifically, the preprocessed data is divided into a training set, a validation set, and a test set, with the ratios being 70%, 15%, and 15%, respectively. The training set is used for model training, the validation set is used for hyperparameter optimization, and the test set is used to evaluate the final performance of the model.

[0088] Specifically, in this embodiment of the present invention, the specific steps of using cross-validation and grid search methods to perform architecture and hyperparameter optimization are as follows:

[0089] Set a series of candidate hyperparameter values, including the number of neurons in the LSTM layer, the learning rate, the batch size, the number of training rounds, etc.; use cross-validation methods (such as k-fold cross-validation) to evaluate the model, dividing the training set into k subsets, where k-1 subsets are used for training and 1 subset is used for validation, repeating this k times, and taking the average validation error as the evaluation metric for model performance; based on the cross-validation results, use a grid search method to search for the optimal hyperparameter combination among the candidate hyperparameter values ​​to ensure that the model can accurately fit historical data and has good generalization ability.

[0090] The improved LSTM model is trained using the optimized hyperparameters, using the mean square error (MSE) as the loss function, and the model parameters are updated through the back propagation algorithm until the model converges.

[0091] Use the trained improved LSTM model to predict renewable energy power generation in the future (e.g., monthly or annually), generating a predicted value for renewable energy power generation at that time. The prediction result should include a specific timestamp and the corresponding predicted power generation value.

[0092] As can be seen from the above description, the embodiments of the present invention provide a method for forecasting renewable energy power generation based on an improved LSTM model. By introducing periodic weights and a memory decay compensation mechanism into the traditional LSTM model, this method fully captures the periodicity and long-term dependency information in historical power generation and meteorological data, thereby improving the accuracy of medium- and long-term forecasts of renewable energy power generation. Furthermore, by extracting spatial features and selecting key features from multidimensional meteorological data, this method effectively reduces redundant information, improves the model's training efficiency and generalization capabilities, and alleviates the technical issues of insufficient forecasting accuracy and poor model stability that exist in traditional forecasting methods when dealing with long-term trends, seasonal changes, and complex meteorological influences.

[0093] Example 2

[0094] Figure 2 Schematic diagram of a new energy power generation prediction system based on an improved LSTM model according to an embodiment of the present invention, which is applied to a new energy power station. Figure 2 As shown, it includes: an acquisition module 10, a preprocessing module 20 and a prediction module 30.

[0095] Specifically, the acquisition module 10 is used to acquire historical power generation data and multi-dimensional meteorological grid data of the new energy power station.

[0096] The preprocessing module 20 is used to preprocess and extract features from historical power generation data and multi-dimensional meteorological grid data to obtain preprocessed data.

[0097] Specifically, the preprocessing module 20 is further configured to:

[0098] Performing data cleaning and normalization on the historical power generation data and the multidimensional meteorological raster data to obtain normalized historical power generation data and normalized multidimensional meteorological raster data respectively;

[0099] Based on the spatiotemporal convolutional network, local spatial features are extracted from the normalized multidimensional meteorological raster data to obtain multidimensional meteorological spatial feature data;

[0100] The multi-dimensional meteorological spatial characteristic data is integrated with the normalized historical power generation data to obtain integrated data;

[0101] The fused data is reduced in dimension and filtered based on the principal component analysis method to obtain the preprocessed data.

[0102] The prediction module 30 is used to use the pre-processed data as input and use the trained improved LSTM model to predict the power generation of the new energy power station to obtain a power generation prediction value.

[0103] Among them, the improved LSTM model includes input gate, forget gate and output gate. The input gate, forget gate and output gate are all periodic memory gating units that introduce periodic weight terms. During the training process of the improved LSTM model, a memory decay compensation mechanism is introduced and cross-validation and grid search methods are used to optimize the architecture and hyperparameters.

[0104] Specifically, if Figure 2 As shown, the system provided by the embodiment of the present invention further includes a training module 40 for training the improved LSTM model to obtain a trained improved LSTM model.

[0105] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method provided in the embodiment of the present invention when executing the computer program.

[0106] The present invention also provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed by a processor, the method provided in the embodiment of the present invention is implemented.

[0107] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0108] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A new energy power generation prediction method based on an improved LSTM model, characterized in that: Applied to new energy power stations; including: Acquiring historical power generation data and multi-dimensional meteorological grid data of the new energy power station; Preprocessing and feature extraction are performed on the historical power generation data and the multidimensional meteorological grid data to obtain preprocessed data; Using the preprocessed data as input, the trained improved LSTM model is used to predict the power generation of the new energy power station to obtain a power generation prediction value; The improved LSTM model includes an input gate, a forget gate, and an output gate, and the input gate, the forget gate, and the output gate are all periodic memory gating units that introduce a periodic weight term. During the training process of the improved LSTM model, a memory decay compensation mechanism is introduced, and cross-validation and grid search methods are used to optimize the architecture and hyperparameters. The structural formula of the input gate in the improved LSTM model includes: ， Where: Indicates Input the activation value of the gate at any moment; Represents the Sigmoid activation function; represents the weight matrix of the input gate; express The hidden state of the moment; express Input data at the moment; represents the bias of the input gate; Represents the periodic weight term introduced in the input gate; Represents a time variable or moment index; The structural formula of the forget gate in the improved LSTM model includes: ， Where: Indicates The activation value of the momentary forget gate; Represents the weight matrix of the forget gate; Represents the bias of the forget gate; Represents the periodic weight term introduced in the forget gate; represents the memory decay compensation term; The structural formula of the output gate in the improved LSTM model includes: ， Where: Indicates Output the activation value of the gate at that moment; Represents the weight matrix of the output gate; Represents the bias of the output gate; represents the periodic weight term introduced in the output gate; The period weight items include: ， in, Indicates The period weight item corresponding to the moment; is a learnable scaling parameter used to adjust the amplitude of the cycle weight during the learning process; Respectively Fourier cosine coefficients and sine coefficients corresponding to the periodic components; represents the fundamental frequency of the nth periodic component; is the index of the periodic component; represents the total number of periodic components considered.

2. The method according to claim 1, wherein: The historical power generation data includes the historical power generation time series data of the new energy power station; The multidimensional meteorological grid data includes wind speed, temperature, humidity, and irradiance.

3. The method according to claim 1, wherein: Preprocessing and feature extraction are performed on the historical power generation data and the multidimensional meteorological grid data to obtain preprocessed data, including: performing data cleaning and normalization processing on the historical power generation data and the multidimensional meteorological grid data to obtain normalized historical power generation data and normalized multidimensional meteorological grid data, respectively; Extracting local spatial features from the normalized multidimensional meteorological raster data based on a spatiotemporal convolutional network to obtain multidimensional meteorological spatial feature data; fusing the multidimensional meteorological spatial feature data with the normalized historical power generation data to obtain fused data; The fused data is subjected to dimensionality reduction and screening based on a principal component analysis method to obtain preprocessed data.

4. The method according to claim 1, wherein: The unit state update formula of the improved LSTM model includes: ， Where: Indicates The unit state at the moment; represents the cell state at time t-1; represents element-wise product; represents the hyperbolic tangent activation function; represents the weight matrix used to calculate the candidate unit state; Indicates the corresponding bias; The hidden state update formula of the improved LSTM model includes: ， Where: express The hidden state of the moment.

5. The method according to claim 1, wherein: The memory decay compensation term is adaptively adjusted according to the second-order derivative information of the loss function of the improved LSTM model with respect to the time step during the training process; The calculation formula of the memory decay compensation term includes: ， Where: Indicates The memory decay compensation term at the moment; is the adjustment coefficient; Represents the loss function used by the improved LSTM model during training.

6. A new energy power generation prediction system based on an improved LSTM model, characterized in that: Applied to new energy power stations; including: acquisition module, preprocessing module and prediction module; among them, The acquisition module is used to acquire the historical power generation data and multi-dimensional meteorological grid data of the new energy power station; The preprocessing module is used to preprocess and extract features from the historical power generation data and the multi-dimensional meteorological grid data to obtain preprocessed data; The prediction module is configured to use the preprocessed data as input and use the trained improved LSTM model to predict the power generation of the new energy power station to obtain a power generation prediction value; The improved LSTM model includes an input gate, a forget gate, and an output gate, and the input gate, the forget gate, and the output gate are all periodic memory gating units that introduce a periodic weight term. During the training process of the improved LSTM model, a memory decay compensation mechanism is introduced, and cross-validation and grid search methods are used to optimize the architecture and hyperparameters. The structural formula of the input gate in the improved LSTM model includes: ， Where: Indicates Input the activation value of the gate at any moment; Represents the Sigmoid activation function; represents the weight matrix of the input gate; express The hidden state of the moment; express Input data at the moment; represents the bias of the input gate; Represents the periodic weight term introduced in the input gate; Represents a time variable or moment index; The structural formula of the forget gate in the improved LSTM model includes: ， Where: Indicates The activation value of the momentary forget gate; Represents the weight matrix of the forget gate; Represents the bias of the forget gate; Represents the periodic weight term introduced in the forget gate; represents the memory decay compensation term; The structural formula of the output gate in the improved LSTM model includes: ， Where: Indicates Output the activation value of the gate at that moment; Represents the weight matrix of the output gate; Represents the bias of the output gate; represents the periodic weight term introduced in the output gate; The period weight items include: ， in, Indicates The period weight item corresponding to the moment; is a learnable scaling parameter used to adjust the amplitude of the cycle weight during the learning process; Respectively Fourier cosine coefficients and sine coefficients corresponding to the periodic components; represents the fundamental frequency of the nth periodic component; is the index of the periodic component; represents the total number of periodic components considered.

7. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images