New energy generation power prediction method based on BiLSTM-TPA-MTCN

Abstract

The invention discloses a new energy power generation power prediction method based on BiLSTM-TPA-MTCN, and relates to the technical field of wind power and photovoltaic power generation power prediction based on multi-feature fusion of a power system. Historical data of a wind power generation system and a photovoltaic power generation system are collected, the historical data are preprocessed to obtain historical collated data, a BiLSTM is utilized to construct a power generation power prediction model, TPA and MTCN are combined to optimize the power generation power prediction model, and the optimized power generation power prediction model is utilized to perform power generation power prediction. The model can provide more accurate wind power and photovoltaic power generation power prediction results when facing complex spatio-temporal data, and the method not only has advantages in short-term prediction, but also provides powerful support for long-term prediction and power grid dispatching optimization, and has great application potential and market prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation prediction for electric power systems, and in particular to a renewable energy power generation prediction method based on a BiLSTM-TPA-MTCN model. Background Art

[0002] Renewable energy power generation prediction is an important means to ensure the safe and stable operation of the power grid. Wind power and photovoltaic power generation have the characteristics of randomness, volatility and intermittency, which makes it challenging to accurately predict their power generation. Accurate prediction of renewable energy power generation can not only improve the dispatch efficiency of the power system, but also reduce the phenomenon of wind and solar power abandonment and promote the consumption of new energy.

[0003] There are many challenges in predicting wind power and photovoltaic power generation. First, wind power is greatly affected by changes in wind speed, and wind speed itself is random and unstable, which makes wind power prediction difficult. Secondly, photovoltaic power generation is significantly affected by weather conditions. Factors such as cloud cover and temperature changes will affect the power generation efficiency of photovoltaic panels. In addition, there are complex spatiotemporal correlations between wind turbines and photovoltaic power stations in different geographical locations, and these factors need to be considered comprehensively to improve prediction accuracy. Although traditional prediction methods such as the ARMA model are simple, they have low accuracy and long calculation time.

[0004] In order to meet these challenges, deep learning technology has been widely used in renewable energy power generation prediction in recent years. In particular, BiLSTM is widely used because it can effectively process time series data. However, the traditional BiLSTM model still has limitations when dealing with the complex relationship between multiple wind power and photovoltaic power stations, and its prediction accuracy and stability need to be improved. Summary of the invention

[0005] The purpose of the present invention is to provide a new energy power generation power prediction method based on BiLSTM-TPA-MTCN to solve the problems raised in the above background technology.

[0006] A new energy power generation power prediction method based on BiLSTM-TPA-MTCN includes the following steps:

[0007] Collect historical data of wind power generation system and photovoltaic power generation system, and pre-process the historical data to obtain historical collated data;

[0008] Power generation prediction model construction: historical data is used as input data, and a bidirectional long short-term memory network model (BiLSTM) is used to perform time series analysis on the historical data to capture the long-term dependencies in the time series data in the historical data, so as to understand the changing trend of power generation and learn the time characteristics of wind power and photovoltaic power generation prediction;

[0009] The first optimization of the power generation prediction model: Combined with the Temporal Pattern Attention (TPA) model, different weights are assigned to each historical collation data, so that the BiLSTM model can focus more on key information, thereby optimizing the BiLSTM model to obtain the BiLSTM-TPA model, and improving the accuracy of wind power and photovoltaic power generation prediction;

[0010] Second optimization of the power generation prediction model: Based on the BiLSTM-TPA model, the multi-head temporal convolutional neural network (MTCN) is used to perform deep learning on the time series features in the BiLSTM-TPA output data to improve the accuracy and precision of the prediction results, so as to mine and capture the multi-scale time features in the historical collated data of wind power and photovoltaic power generation;

[0011] The optimized power generation prediction model is used to predict power generation: the current real-time data of wind power and photovoltaic power generation are provided as input data to the optimized power generation prediction model, and the optimized power generation prediction model outputs the corresponding wind power and photovoltaic power generation power prediction values, and the output wind power and photovoltaic power generation power prediction values ​​are used as the power generation prediction results within a predetermined time range in the future.

[0012] In the above-mentioned new energy power generation prediction method based on BiLSTM-TPA-MTCN, a power generation prediction model is constructed based on BiLSTM, and TPA and MTCN are used to jointly optimize the constructed power generation prediction model, and finally an optimized power generation prediction model is generated. This model can provide more accurate wind power and photovoltaic power generation prediction results when facing complex spatiotemporal data. This method not only has advantages in short-term prediction, but also provides strong support for long-term prediction and power grid dispatch optimization, and has great application potential and market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a process for predicting renewable energy power generation based on BiLSTM-TPA-MTCN in an embodiment

[0014] Figure 2 It is a schematic diagram of the BiLSTM structure;

[0015] Figure 3 It is a schematic diagram of the TPA structure;

[0016] Figure 4 Schematic diagram of the TCN residual structure. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the embodiment of the present invention provides a new energy power generation power prediction method based on BiLSTM-TPA-MTCN, comprising the following steps:

[0019] Step S1, collect historical data of wind power generation system and photovoltaic power generation system, and pre-process the historical data to obtain historical collated data. The historical collated data includes meteorological data, temperature data, humidity data, wind speed data, irradiance data, historical wind power generation data, photovoltaic power generation data and time data corresponding to the above data; among them, wind speed data is the main factor affecting wind power; photovoltaic data is the main factor affecting photovoltaic power, and the data is obtained from satellites or ground monitoring stations, and its time resolution is ensured to be consistent with photovoltaic power data; meteorological data, temperature data, and humidity data will directly affect wind power generation, and will also affect the efficiency of photovoltaic panels. Meteorological data, temperature data, and humidity data need to be obtained from meteorological stations, appropriately processed, and their time resolution is ensured to be consistent with wind power data. During the data collection process, wind power data needs to be collected at wind speed sensors at different heights (such as 10 meters, 30 meters, 50 meters, 70 meters, etc.), and the hub height, wind speed, wind direction, temperature, air pressure, humidity, time and other information of the wind turbine are recorded at the same time. Photovoltaic data requires information such as total radiation, normal direct irradiance, horizontal plane scattered irradiance, temperature, air pressure, time and humidity. In the data preprocessing stage, the collected multidimensional data is processed by normalization or standardization methods to adapt it to the deep neural network model. In particular, different processing strategies are used for data of different feature dimensions to reduce the impact of noise on model training and ensure the quality and accuracy of the data.

[0020] Among them, the preprocessing of historical data is specifically as follows: historical data cleaning and normalization processing. First, the correlation analysis method is used to screen out the influencing factors that are highly correlated with wind power and photovoltaic power generation. Since different sensors may have measurement errors, the screened historical data needs to be cleaned to remove outliers, noise, duplicate data and missing values. For missing data, interpolation or filling methods based on surrounding data are used for processing. At the same time, all data are normalized to the same scale to facilitate subsequent model training. The processed historical data is divided into three categories, namely training set, validation set and test set. 70%-80% of the processed historical data is used for training, 10%-15% of the processed historical data is used for validation, and the rest is used for testing.

[0021] Step S3, constructing a power generation prediction model: using historical data (e.g., using the above training set) as input data, and using a bidirectional long short-term memory network (BiLSTM) to perform time series analysis on the historical data to capture the long-term dependencies in the time series data in the historical data, that is, the correlation or influence between the data points in the time series and the data at the previous distant time points, so as to understand the changing trend of power generation and learn the time characteristics of wind power and photovoltaic power generation prediction; please also refer to Figure 2 BiLSTM is an improved recurrent neural network (RNN) that can consider both past and future information in sequence data. The BiLSTM model includes two parts: forward LSTM and reverse LSTM, which respectively extract forward and backward time series information of input data to improve the accuracy of power prediction. The BiLSTM model uses a bidirectional long short-term memory unit (LSTM) to perform bidirectional time series modeling on historical data, learn long-term dependencies and short-term change rules in the data, and then learn how to predict future power generation based on past power change trends, meteorological data and other environmental factors.

[0022] Step S5, first optimization of the power generation prediction model: Combined with the Temporal Pattern Attention (TPA) model, different weights are assigned to each historical collation data (for example, using the above training set), so that the BiLSTM model can focus more on key information, thereby optimizing the BiLSTM model to obtain the BiLSTM-TPA model and improve the accuracy of wind power and photovoltaic power generation prediction; add the TPA mechanism to the BiLSTM, and extract the complex connections between different wind turbines or photovoltaic power stations by analyzing the state vector of the BiLSTM hidden layer, and select relevant information from different time scales. This mechanism can identify information at different time steps and select relevant information according to the characteristics of each time step, thereby reducing prediction errors. Please also refer to Figure 3 ,The TPA mechanism is particularly suitable for dealing with the ,power prediction problems of multiple wind turbines or photovoltaic power plants, because TPA focuses on the ,temporal characteristics of different time steps, and highlights the key ,information through weighted allocation, and appropriately adapts to non-stationary load data, and it can ,simultaneously consider the interdependencies among multiple variables, i.e., multiple factors affecting the ,wind power or photovoltaic power generation.

[0023] Step S7, second optimization of power generation prediction model: Based on the BiLSTM-TPA model, the multi-head temporal convolutional neural network (MTCN) is used to further perform deep learning on the time series features in the BiLSTM-TPA output data to improve the accuracy and precision of the prediction results, so as to mine and capture the multi-scale time features in the historical collated data of wind power and photovoltaic power generation, such as the trend of wind speed and photovoltaic radiation changing with the seasons; wind power generation has stronger winds in some seasons, and photovoltaic power generation has higher power generation in seasons with sufficient sunshine; the changes in wind speed and photovoltaic radiation in different time periods of the day, the correlation between data at different time points and other features; Among them, the output data of BiLSTM-TPA is obtained based on the validation set and test set of the historical collated data; MTCN is an extension based on TCN, please also refer to Figure 4, where TCN includes a stack of convolutions, causal convolutions, dilated convolutions, and residual connections. It can process multiple time series or multiple input channels at the same time. MTCN processes data of different time scales through multiple parallel TCNs, and deeply learns the time series features in the data, such as the trend of wind speed and photovoltaic radiation changing with the seasons; wind power generation has stronger winds in some seasons, and photovoltaic power generation has higher power generation in seasons with sufficient sunshine; the changes in wind speed and photovoltaic radiation in different time periods of the day, the correlation between data at different time points, and other characteristics, and then fuses these features of different scales to capture the complex patterns in the time series data and perform error correction on the prediction model to improve the final prediction accuracy.

[0024] Step S9, use the optimized power generation prediction model to predict power generation: the acquired wind speed, temperature, total radiation, power generation, etc. of wind power and photovoltaic power generation are provided as input data to the optimized power generation prediction model, and the optimized power generation prediction model outputs the corresponding wind power and photovoltaic power generation power prediction values, and the output wind power and photovoltaic power generation power prediction values ​​are used as the power generation prediction results within a predetermined time range in the future.

[0025] In order to make the technical means, creative features, objectives and effects achieved by this application easy to understand, the technical means in the technical solution of this application are introduced as follows:

[0026] 1. Preprocess the historical data of wind power generation system and photovoltaic power generation system to obtain historical collated data:

[0027] (1) Data cleaning

[0028] The collected historical data may contain missing values, outliers, or duplicate data, which need to be cleaned. The steps for cleaning historical data include:

[0029] Missing value processing: For missing data, linear interpolation, mean filling or nearest neighbor filling can be used to fill in the missing data points.

[0030] Outlier processing: Detect outliers in the data through statistical methods (such as Z-Score) or box plot method. If abnormal data points are found, they can be repaired by median replacement, outlier deletion or smoothing.

[0031] Deduplication: Checks the data for duplicate timestamps. If duplicate data points are found, retains the only data record and removes the redundant data.

[0032] (2) Time series data division

[0033] Renewable energy power generation prediction is based on time series data, so historical data needs to be divided in chronological order to input effective time steps when training the model. The historical data division includes:

[0034] Training set: Usually 60%-80% of the data is selected as the training set for model training and parameter adjustment.

[0035] Validation set: Select 10%-20% of the data as the validation set for model tuning to prevent overfitting.

[0036] Test set: Select the remaining data as the test set for the final evaluation of the model’s performance.

[0037] (3) Standardization and normalization of historical data

[0038] In order to avoid the negative impact of inconsistent input data dimensions on model training, the data needs to be standardized or normalized. Common methods include Z-score standardization and Min-Max normalization. Z-score standardization: convert the data into a distribution with a mean of 0 and a standard deviation of 1:

[0039]

[0040] Among them, X is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0041] Min-Max Normalization: Scale the data to a specified interval (usually [0,1]):

[0042]

[0043] Among them, X min ,X max are the minimum and maximum values ​​of the data respectively.

[0044] (4) Historical data segmentation and time series input

[0045] Since the prediction of power generation is based on time series data, historical data needs to be divided into multiple subsequences in chronological order. Each subsequence contains a certain number of time steps as input features, and the prediction target is the power generation in one or more future time steps. Suppose we use T time steps of data to predict future power, where each time step has m features, the dimension of the input data is (n×m), where n is the number of subsequences into which the data is divided by time, and the dimension of the output data is (1×k), where k is the number of predicted power generation steps.

[0046] (5) Sequence Sliding Window

[0047] To generate training data, a sliding window method is used. For a given time series {P1,P2,...,P T}, the time sliding window size is ω duration, the prediction duration is h, then the training samples extracted after each sliding are:

[0048] (X t ,y t+h )whereX t = {P t-ω+1 ,P t-ω+2 ,...,P t},y t+h =P t+h

[0049] (6) Data format conversion

[0050] Finally, the processed data is converted into the format required for model input to form historical collated data, and finally input into the BiLSTM model for training.

[0051] 2. Construction of power generation prediction model:

[0052] BiLSTM (bidirectional long short-term memory network model) is used to perform time series analysis on the historical data to capture the long-term dependencies in the time series data in the historical data, that is, the correlation or influence between the data points in the time series and the data at the previous distant time points, so as to understand the changing trend of power generation and learn the time characteristics of wind power and photovoltaic power generation prediction. The BiLSTM model captures past and future information simultaneously through two LSTM layers, forward and reverse, so as to effectively capture the long-term dependencies of the time series. Assume that the input historical data sequence of wind speed, temperature, etc. is X = [x1, x2, ..., x T ]. The processing of BiLSTM is as follows:

[0053] h tf =LSTM f (x t ,h t-1,f )

[0054] h tb =LSTM b (x t ,h t+1,b )

[0055] h t =concat[h tf ,h tb ]

[0056] x t Represents the current input of the time series;

[0057] h t Represents the hidden state vector of BiLSTM;

[0058] concat means concatenation operation in the output dimension;

[0059] h tf ,h tb Represent the hidden state vectors of the forward and backward LSTM respectively.

[0060] 3. First optimization of power generation prediction model:

[0061] Based on the BiLSTM network model, because renewable energy generation (such as wind power and photovoltaics) is affected by a variety of time-related factors, such as wind speed, light intensity, weather conditions, etc., adding the TPA mechanism can provide a mechanism for the model to identify and emphasize the information at these key time points, rather than just averaging the data at all time points. It can also help the model capture these dynamic changes. By assigning different weights to data at different time points, the model can better understand the changing trend of power generation over time.

[0062] (1) CNN detection of temporal pattern

[0063] TPA uses a multi-dimensional CNN filter to extract significant features from the BiLSTM hidden state matrix. First, use BiLSTM to process the input data (for example, using the above training set) to obtain the hidden state vector h for each timestep i . Among them, ω is the time window length, representing the length of the selected input sequence, and the hidden state matrix H = h t-ω ,h t-ω+1 ,...,h t-1 .

[0064] In the hidden state matrix H, one-dimensional convolution is used to convolve along the m features of H to extract the relationship between features such as temperature, humidity, wind speed, radiation, power, and the time pattern matrix H. C , C is the filter,

[0065]

[0066] in, is the convolution value of the i-th row vector and the j-th filter. i,(t-ω-1+l) In it, t is the time, ω is the time window length, and l is the padding length in the convolution calculation. * is the convolution calculation performed, and T represents the maximum length of the weight extracted by the filter.

[0067] (2) Attention Weighting

[0068] 1) Define f as an evaluation function that determines the relevance by calculating the weight, and the result of the weight calculation is as follows:

[0069]

[0070] 2) Use sigmoid for normalization to obtain attention weights, which facilitates the selection of multiple variables:

[0071]

[0072] 3) Using attention weights, H C Do a weighted sum on each row to get the attention vector v t :

[0073]

[0074] m represents the number of features of the input variable;

[0075] v t With h t The predicted power generation is obtained by linear mapping and adding:

[0076] y t-1+Δ =W h' h' t =W h' (W h h t +W v v t )

[0077] y t-1+Δ Indicates the predicted power generation value; h t is an intermediate variable used to generate the final value; Δ represents the prediction time scale of different prediction tasks; W h' ,W h and W v are different weight matrices corresponding to the variables.

[0078] 4. Second optimization of power generation prediction model:

[0079] The core of TCN is the use of dilated convolution. This convolution structure can increase the receptive field of the network without increasing the number of parameters, thereby effectively capturing long-term dependencies.

[0080] The subsequence is predicted through the BiLSTM-TPA model, and the subsequence prediction results are reconstructed. After the prediction results are obtained, the prediction error is obtained based on the prediction results and the historical actual power generation corresponding to the input data of the prediction results. The prediction error and historical collated data such as temperature, humidity, wind speed, and radiation are used as input feature sequences. TCN can use these data to help correct the prediction error, thereby improving the accuracy of the final prediction results. We have an input sequence x1, x2, ..., x t , and hope to predict the corresponding output y1,y2,...,y at each time point t The key constraint is that when predicting the output at a certain point in time, the previously observed input can be used to predict the current value y in the causal convolution. t The generation of t and the previous input, but not the future input x t+1 Related. It can be expressed as:

[0081] y0,...y t =f(x0,...x t )

[0082] (1) Causal Convolution

[0083] TCN uses causal convolution to ensure that the model does not violate the time order. For causal convolution, the output y(t) only depends on the current and past time steps, that is:

[0084]

[0085] ω is the convolution kernel; k is the convolution kernel size; x is the input sequence.

[0086] (2) Dilated Convolution

[0087] In order to increase the receptive field without increasing the number of parameters, TCN dilated convolution is used to capture a larger receptive field without increasing the number of parameters, thereby better understanding the contextual information in the input data.

[0088]

[0089] d is the expansion rate;

[0090] (3) Residual Connection

[0091] In deep networks, directly connecting input and output can avoid the gradient vanishing problem and help converge faster. After each convolutional layer, a residual connection is usually added to maintain information flow. Assuming the output of the convolutional layer is h(t), the output after the residual connection is:

[0092] y(t)=h(t)+x(t)

[0093] x(t) is the signal input to this layer; h(t) is the output of the convolutional layer.

[0094] (4) Output layer:

[0095] Finally, the model maps the output of the convolutional layer to the final power generation prediction value through the fully connected layer. In the regression task, the output It can be obtained by the following formula:

[0096]

[0097] W is the weight matrix of the convolutional layer; y(t) is the feature extracted by the convolutional layer; b is the bias term.

[0098] MTCN processes data of different time scales through multiple parallel TCNs, and then fuses the features of these different scales to capture the complex patterns in time series data. Compared with traditional TCN, MTCN has a larger effective receptive field (ERFs), so it can better exert the potential of convolution in time series analysis.

[0099] 5. Use the optimized power generation prediction model to predict power generation:

[0100] (1) Model training

[0101] The optimized power generation prediction model (BiLSTM-TPA-MTCN model) is trained using historical collated data. During the training process, the cross-validation method can be used to optimize the model parameters, and the mean square error (MSE) is used as the loss function:

[0102]

[0103] where y i is the actual value; is the predicted value; n is the number of samples;

[0104] (2) Power generation prediction

[0105] After the training is completed, the trained model is used for real-time power prediction. Input real-time wind speed, temperature, humidity and other data, and obtain the final power generation prediction result through the optimized power generation prediction model:

[0106] P pred =f(X input )

[0107] P pred is the predicted power; X input is the input feature; f is the BiLSTM-TPA-MTCN model;

[0108] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A new energy power generation power prediction method based on BiLSTM-TPA-MTCN, comprising the following steps: Collect historical data of wind power generation system and photovoltaic power generation system, and pre-process the historical data to obtain historical collated data; Power generation prediction model construction: historical data is used as input data, and BiLSTM is used to perform time series analysis on the historical data to capture the long-term dependencies in the time series data in the historical data, so as to understand the changing trend of power generation and learn the time characteristics of wind power and photovoltaic power generation prediction; The first optimization of the power generation prediction model: Combined with the TPA model, different weights are assigned to each historical collation data, so that the BiLSTM model can focus more on key information, thereby optimizing the BiLSTM model to obtain the BiLSTM-TPA model, and improving the accuracy of wind power and photovoltaic power generation prediction; Second optimization of the power generation prediction model: Based on the BiLSTM-TPA model, MTCN is used to perform deep learning on the time series features in the BiLSTM-TPA output data to improve the accuracy and precision of the prediction results, so as to mine and capture the multi-scale time features in the historical collated data of wind power and photovoltaic power generation; The optimized power generation prediction model is used to predict power generation: the current real-time data of wind power and photovoltaic power generation are provided as input data to the optimized power generation prediction model, and the optimized power generation prediction model outputs the corresponding wind power and photovoltaic power generation power prediction values, and the output wind power and photovoltaic power generation power prediction values ​​are used as the power generation prediction results within a predetermined time range in the future.

2. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 1 is characterized in that: The historical collated data includes meteorological data, temperature data, humidity data, wind speed data, irradiance data, historical wind power generation data, photovoltaic power generation data and the time data corresponding to the above data.

3. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 2 is characterized in that: Preprocess the historical data, including: Firstly, the correlation analysis method is used to screen out the influencing factors that are highly correlated with wind power and photovoltaic power generation; Secondly, interpolation or filling based on surrounding data is used for processing, and all data are normalized to the same scale to facilitate subsequent model training; Finally, the processed historical data is divided into three categories, namely training set, validation set and test set. 70%-80% of the processed historical data is used for training, 10%-15% of the processed historical data is used for validation, and the rest is used for testing.

4. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 1 is characterized in that: The process of building the power generation prediction model is as follows: The sequence of historical data input is X = [x1, x2, ..., x T ], the processing of BiLSTM is as follows: h tf =LSTM f (x t ,h t-1,f ) h tb =LSTM b (x t ,h t+1,b ) h t =concat[h tf ,h tb ] x t Represents the current input of the time series; h t Represents the hidden state vector of BiLSTM; concat means concatenation operation in the output dimension; h tf ,h tb They represent the hidden state vectors of the forward and backward LSTMs respectively, and the hidden state vectors are the long-term dependencies.

5. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 4 is characterized in that: The first optimization process of the power generation prediction model is: TPA uses a multi-dimensional CNN filter to extract significant features from the BiLSTM hidden state matrix. First, the input data is processed using BiLSTM to obtain the hidden state vector h for each timestep. i , where ω is the time window length, representing the length of the selected input sequence, and the hidden state matrix H = h t-ω ,h t-ω+1 ,...,h t-1 ; In the hidden state matrix H, one-dimensional convolution is used to convolve along the m features of H to extract the relationship between features such as temperature, humidity, wind speed, radiation, power, and the time pattern matrix H. C , C is the filter, in, is the convolution value of the i-th row vector and the j-th filter, i in H i,(t-ω-1+l) In it represents a row vector, t is time, ω is the time window length, l is the padding length in the convolution calculation, * is the convolution calculation performed, and T represents the maximum length of the weight extracted by the filter; (2) Attention Weighting 1) Define f as an evaluation function that determines relevance by calculating weights, and the result of the weight calculation is as follows: 2) Use sigmoid for normalization to obtain attention weights, which facilitates the selection of multiple variables: 3) Using attention weights, H C Do a weighted sum on each row to get the attention vector v t : m represents the number of features of the input variable; v t With h t The predicted power generation is obtained by linear mapping and adding: y t-1+Δ =W h' h' t =W h' (W h h t +W v v t ) y t-1+Δ Indicates the predicted power generation value; h t is an intermediate variable used to generate the final value; Δ represents the prediction time scale of different prediction tasks; W h' ,W h and W v are different weight matrices corresponding to the variables.

6. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 5 is characterized in that: The second optimization process of the power generation prediction model is: The subsequence is predicted by the BiLSTM-TPA model, and the subsequence prediction results are reconstructed. After the prediction results are obtained, the prediction error is obtained based on the prediction results and the historical actual power generation corresponding to the input data of the prediction results. The prediction error and historical collated data such as temperature, humidity, wind speed, and radiation are used as input feature sequences. TCN can use these data to help correct the prediction error, thereby improving the accuracy of the final prediction result; specifically: the input sequence x1, x2, ..., x t , and hope to predict the corresponding output y1,y2,...,y at each time point t ; The current prediction value y in the causal convolution t The generation of t and the previous input, but not the future input x t+1 Related, it can be expressed as: y0,...y t =f(x0,...x t ) (1) Causal Convolution TCN uses causal convolution to ensure that the model does not violate the time order; for causal convolution, the output y(t) only depends on the current and past time steps, that is: ω is the convolution kernel; k is the convolution kernel size; x is the input sequence; (2) Dilated Convolution In order to increase the receptive field without increasing the number of parameters, TCN dilated convolution is used to capture a larger receptive field without increasing the number of parameters, thereby better understanding the contextual information in the input data; d is the expansion rate; (3) Residual Connection In deep networks, directly connecting input and output can avoid the gradient vanishing problem and help converge faster. After each convolutional layer, a residual connection is usually added to maintain information flow. The output of the convolutional layer is h(t), and the output after the residual connection is: y(t)=h(t)+x(t) x(t) is the signal input to this layer; h(t) is the output of the convolutional layer; (4) Output layer: Finally, the model maps the output of the convolutional layer to the final power generation prediction value through the fully connected layer. In the regression task, the output It can be obtained by the following formula: W is the weight matrix of the convolutional layer; y(t) is the feature extracted by the convolutional layer; b is the bias term.

7. The new energy power generation prediction method based on BiLSTM-TPA-MTCN according to claim 1 is characterized in that: The process of power generation prediction using the optimized power generation prediction model is as follows: The optimized power generation prediction model is trained using historical collated data. During the training process, the cross-validation method can be used to optimize the model parameters, and the mean square error (MSE) is used as the loss function: where y i is the actual value; is the predicted value; n is the number of samples; After the training is completed, the trained model is used for real-time power prediction: the current real-time data of wind power and photovoltaic power generation is input, and the final power generation prediction result is obtained through the optimized power generation prediction model: P pred =f(X input ) P pred is the predicted power; X input is the input feature; f is the optimized power generation prediction model.

Images