Short-term wind and light output prediction method based on double-branch deep learning model
By using a dual-branch deep learning model to extract and fuse features from meteorological time series and spatial images, the problem of missing spatial information and noise interference in wind and solar power output prediction in existing technologies is solved, and more accurate wind and solar power output prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-10
AI Technical Summary
Existing short-term wind and solar power output forecasting methods suffer from spatial information loss and noise interference when using multi-source heterogeneous data, resulting in high model learning difficulty, poor generalization ability, and difficulty in accurately predicting power output changes caused by complex weather phenomena.
A dual-branch deep learning model is adopted. By performing wavelet transform processing on meteorological time series to remove noise and extract temporal features, and combining it with meteorological spatial images to extract spatial features, CNN and LSTM networks are used to extract features respectively and then fused to generate more accurate wind and solar power output prediction results.
The model's robustness and generalization ability have been improved, enabling it to more accurately capture rapid changes in output caused by complex space weather phenomena, thus enhancing prediction accuracy and stability.
Smart Images
Figure CN121840573A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a wind-solar power output prediction method, in particular to a short-term wind-solar power output prediction method based on a double-branch deep learning model. BACKGROUND
[0002] In the prior art, short-term wind-solar power output prediction methods can be mainly divided into methods based on physical models, traditional statistical methods, and data-driven machine learning / deep learning methods. However, these prior arts generally have the following core defects and deficiencies when facing complex practical application scenarios:
[0003] Traditional statistical methods and early machine learning models mainly rely on single-point observation time series data of meteorological stations (such as wind speed and irradiance). Although such data contains time evolution information, it completely loses the spatial structure information of clouds, wind fields, etc. Meteorological image data such as satellite cloud images and radar echo images contain key spatial features (such as cloud shape, coverage area, and moving track) that affect wind-solar power output, resulting in insufficient use of multi-source heterogeneous data, loss of spatial information, and slow response to power output sudden rise or drop events caused by local weather mutations (such as rapidly moving clouds and sudden gusts).
[0004] Since the original time series signals of wind-solar power output and related meteorological elements have significant non-stationarity, randomness, and strong noise interference. Directly inputting such original signals into a prediction model (such as LSTM and GRU) requires the model not only to learn the complex output law, but also to "passively" handle noise and trend fluctuations, which greatly increases the learning difficulty, leading to slow model convergence and poor generalization ability. The model is prone to overfitting to noise in the training data, and performs poorly on unseen fluctuation patterns, making it difficult to guarantee the stability and robustness of the prediction results.
[0005] Therefore, in order to solve the above technical problems, it is necessary to propose a new technical means. SUMMARY
[0006] Therefore, the purpose of the present application is to provide a short-term wind-solar power output prediction method based on a double-branch deep learning model, which acquires meteorological time series, meteorological spatial images, and historical output, then performs wavelet transform processing on the meteorological time series to effectively remove data noise and smooth the signal, then acquires time scale features and spatial scale features from the meteorological time series and meteorological spatial images based on a double-branch deep learning model, and then performs fusion, which can more accurately capture rapid changes in output caused by complex spatial weather phenomena, enhance robustness and generalization ability, and effectively ensure the final prediction accuracy.
[0007] The application provides a short-term wind-solar power output prediction method based on a double-branch deep learning model.
[0008] S1. Obtain wind-solar power generation operation parameters, including meteorological time series data, meteorological spatial image data and historical wind-solar power output data;
[0009] S2. Preprocess the wind-solar power generation operation parameters to form training data;
[0010] S3. Construct a double-branch deep learning model, wherein the double-branch deep learning model comprises a CNN spatial feature extraction network, an LSTM time feature extraction network, a feature fusion module and a prediction output network;
[0011] The preprocessed meteorological spatial image data is input into the CNN spatial feature extraction network, and the preprocessed meteorological time series data is input into the LSTM time feature extraction network; the output features of the CNN spatial feature extraction network and the LSTM time feature extraction network are input into the feature fusion module, the output features of the feature fusion module are input into the prediction output network, and the prediction output network outputs wind-solar power output prediction results;
[0012] S4. Input the preprocessed wind-solar power generation operation parameters into the double-branch deep learning model for training;
[0013] S5. Obtain real-time wind-solar operation data and input the real-time wind-solar operation data into the trained double-branch deep learning model to obtain wind-solar power output prediction results.
[0014] Further, in step S2, preprocessing the meteorological time series data specifically comprises:
[0015] Obtain environmental parameters, including wind speed, wind direction, solar irradiance, air temperature and relative humidity;
[0016] Sample the environmental parameters according to a set time frequency, and arrange the sampled environmental parameters according to the sampling time sequence to form meteorological time series data;
[0017] Interpolate missing points in the meteorological time series data;
[0018] Remove outliers in the meteorological time series data;
[0019] Normalize the meteorological time series data;
[0020] Perform wavelet transform decomposition on the normalized meteorological time series data to obtain training data of the meteorological time series data.
[0021] Further, preprocessing the meteorological spatial image data comprises:
[0022] The meteorological spatial image data is a satellite cloud image or a weather radar echo image;
[0023] The meteorological spatial image data and meteorological time series data are time-stamp aligned to ensure consistency between the meteorological time series data and meteorological spatial image data on a time scale.
[0024] Crop meteorological spatial image data;
[0025] The cropped meteorological spatial image data is resized and unified to a set size;
[0026] The pixel values of the scaled meteorological spatial image data are normalized.
[0027] Constructing image sequence tensors: Where: T is the time step, H, W and C are the height, width and number of channels of the image, respectively; the image sequence tensor is used as the preprocessed meteorological spatial image data for training the two-branch deep learning model.
[0028] Furthermore, the prediction output network includes a regression prediction network composed of fully connected layers and a wavelet inverse transform module;
[0029] The regression prediction network outputs wavelet component features corresponding to the wind and solar power output, and the wavelet inverse transform module performs an inverse transform on the wavelet component features to obtain the final wind and solar power output result.
[0030] Furthermore, the calculation process of the LSTM temporal feature extraction network is as follows:
[0031] Forget Gate Calculation:
[0032] Input gate calculation:
[0033] Candidate cell state calculation:
[0034] Cell status update:
[0035] Output gate calculation:
[0036] Hidden state output: ;
[0037] in: Let the input vector be the input vector at the current time. This is the hidden state from the previous moment. The sigmoid activation function compresses the gated signal into the (0,1) interval, and ⊙ represents element-wise multiplication.
[0038] The beneficial effects of this invention are as follows: By acquiring meteorological time series, meteorological spatial images, and historical power output, and then performing wavelet transform processing on the meteorological time series, data noise is effectively removed and the signal is stabilized. Then, based on a dual-branch deep learning model, temporal and spatial scale features are obtained from the meteorological time series and meteorological spatial images respectively, and then fused. This can more accurately capture the rapid changes in power output caused by complex spatial weather phenomena, enhance robustness and generalization ability, and effectively ensure the accuracy of the final prediction. Attached Figure Description
[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0040] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0041] The present invention will be further described in detail below:
[0042] This invention provides a short-term wind and solar power output prediction method based on a dual-branch deep learning model, comprising the following steps:
[0043] S1. Obtain wind and solar power generation operation parameters, including meteorological time series data, meteorological spatial image data, and historical wind and solar power output data;
[0044] S2. Preprocess the operating parameters of wind and solar power generation to form training data;
[0045] S3. Construct a dual-branch deep learning model, which includes a CNN spatial feature extraction network, an LSTM temporal feature extraction network, a feature fusion module, and a prediction output network;
[0046] The preprocessed meteorological spatial image data is input into the CNN spatial feature extraction network, and the preprocessed meteorological time series data is input into the LSTM temporal feature extraction network. The output features of the CNN spatial feature extraction network and the LSTM temporal feature extraction network are input into the feature fusion module, and the output features of the feature fusion module are input into the prediction output network. The prediction output network outputs the wind and solar power output prediction results.
[0047] The CNN spatial feature extraction network uses existing convolutional neural networks, which consists of convolutional modules, activation modules, pooling modules, flattening modules, and fully connected modules. Its calculation process is existing technology and will not be described in detail here.
[0048] The LSTM temporal feature extraction network adopts the existing long short-term memory network, which effectively captures long-term dependencies through a precise gating mechanism (input gate i_t, forget gate f_t, output gate o_t) and cell state c_t, overcoming the gradient vanishing / exploding problem of traditional recurrent neural networks.
[0049] The calculation process of the LSTM temporal feature extraction network is as follows:
[0050] Forget Gate Calculation:
[0051] Input gate calculation:
[0052] Candidate cell state calculation:
[0053] Cell status update:
[0054] Output gate calculation:
[0055] Hidden state output: ;
[0056] in: Let the input vector be the input vector at the current time. This is the hidden state from the previous moment. The sigmoid activation function compresses the gated signal into the (0,1) interval, and ⊙ represents element-wise multiplication.
[0057] The feature fusion module splices the CNN spatial feature extraction network and the LSTM temporal feature extraction network along the channel direction to form a fused feature. It is performed using an existing splicing module. This operation retains all feature information from the two branches, providing a joint representation that includes spatial context and temporal evolution rules for subsequent prediction, thereby effectively improving the final prediction accuracy.
[0058] S4. Input the preprocessed wind and solar power generation operation parameters into the dual-branch deep learning model for training;
[0059] S5. Acquire real-time wind and solar power operation data and input it into the trained bi-branch deep learning model to obtain the wind and solar power output prediction results. Using the above method, by acquiring meteorological time series, meteorological spatial images, and historical power output data, and then performing wavelet transform processing on the meteorological time series to effectively remove data noise and stabilize the signal, the bi-branch deep learning model extracts temporal and spatial scale features from the meteorological time series and meteorological spatial images respectively, and then fuses them. This allows for more accurate capture of rapid power output changes caused by complex spatial weather phenomena, enhancing robustness and generalization ability, and effectively ensuring the final prediction accuracy.
[0060] Of course, historical output data also needs to be preprocessed, such as removing outliers, interpolating and filling missing values, and normalizing. Historical output data should be strictly consistent with meteorological time series data and meteorological spatial image data in terms of time scale. Historical output data is used as input to the prediction model for comparative loss, thereby updating the network parameters. The two-branch deep learning model uses existing loss functions, such as mean loss and entropy loss, which will not be elaborated here.
[0061] In this embodiment, step S2, the preprocessing of meteorological time series data specifically includes:
[0062] Obtain environmental parameters, including wind speed, wind direction, solar irradiance, air temperature, and relative humidity;
[0063] Environmental parameters are sampled at a set time frequency, and the sampled environmental parameters are arranged according to the sampling time sequence to form meteorological time series data;
[0064] Perform missing point interpolation on meteorological time series data;
[0065] Remove outliers from meteorological time series data;
[0066] Normalize the meteorological time series data;
[0067] The training data for the meteorological time series data is obtained by performing wavelet transform decomposition on the normalized meteorological time series data.
[0068] Among them, wavelet transform is an existing technique that decomposes a signal using wavelet decomposition, thereby analyzing different frequency components of the signal by scaling and shifting the mother wavelet. After layer decomposition (usually) =3 or 4), the original non-stationary sequence Decomposed into an approximate component and Each detail (j=1,2,...,J), its mathematical expression is:
[0069] ;
[0070] It is obtained through a low-pass filter, which characterizes the macroscopic low-frequency trend and main outline of the signal; These signals are obtained through high-pass filters of different scales, representing high-frequency fluctuations, abrupt changes, and noise information at different resolutions (from coarse to fine). This process transforms the original signal, mixed with noise and non-stationary trends, into a set of relatively stationary subsequences with clearer physical meaning. The higher stationarity and more concentrated features of each component sequence greatly reduce the difficulty for subsequent models to directly learn the complex original signal and provide the model with a feature view at multiple time scales.
[0071] In this embodiment, the preprocessing of meteorological spatial image data includes:
[0072] The meteorological spatial image data is a satellite cloud image or a weather radar echo image;
[0073] The meteorological spatial image data and meteorological time series data are time-stamp aligned to ensure consistency between the meteorological time series data and meteorological spatial image data on a time scale.
[0074] Crop meteorological spatial image data;
[0075] The cropped meteorological spatial image data is resized and unified to a set size;
[0076] The pixel values of the scaled meteorological spatial image data are normalized.
[0077] Constructing image sequence tensors: Where: T is the time step, H, W and C are the height, width and number of channels of the image, respectively; the image sequence tensor is used as the preprocessed meteorological spatial image data for training the two-branch deep learning model.
[0078] In this embodiment, the prediction output network includes a regression prediction network composed of fully connected layers and a wavelet inverse transform module;
[0079] The regression prediction network outputs wavelet component features corresponding to the wind and solar power output. The inverse wavelet transform module performs an inverse transform on the wavelet component features to obtain the final wind and solar power output result. Specifically, the fully connected layer outputs wavelet component values corresponding to the wind and solar power output at the future target time (t+Δt). Then, the inverse wavelet transform module performs an inverse transform operation on the output wavelet component prediction values. Through wavelet reconstruction, the model generates a smooth prediction value that conforms to the time-frequency characteristics of the signal, effectively avoiding the problems of unsmooth prediction curves or non-compliance with physical laws that may result from directly regressing the original power output value.
[0080] The inverse wavelet transform (i.e., wavelet reconstruction) used in the inverse wavelet transform module is an existing method and will not be elaborated here.
[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1.A short-term wind-solar power output prediction method based on a two-branch deep learning model, characterized in that: The method comprises the following steps: S1. Obtain wind and light power generation operation parameters, including meteorological time series data, meteorological spatial image data, and historical wind and light output data; S2. Preprocess the wind and light power generation operation parameters to form training data; S3. Construct a double-branch deep learning model, which comprises a CNN spatial feature extraction network, an LSTM time feature extraction network, a feature fusion module, and a prediction output network; The preprocessed meteorological spatial image data is input into the CNN spatial feature extraction network, and the preprocessed meteorological time series data is input into the LSTM time feature extraction network; the output features of the CNN spatial feature extraction network and the LSTM time feature extraction network are input into the feature fusion module, and the output features of the feature fusion module are input into the prediction output network, which outputs the wind and light output prediction result; S4. Input the preprocessed wind and light power generation operation parameters into the double-branch deep learning model for training; S5. Obtain real-time wind and light operation data and input them into the trained double-branch deep learning model to obtain the wind and light output prediction result. 2.The short-term wind-solar power output prediction method based on the dual-branch deep learning model according to claim 1, characterized in that: In step S2, the preprocessing of the meteorological time series data specifically comprises: Obtain environmental parameters, including wind speed, wind direction, solar irradiance, air temperature, and relative humidity; Sample the environmental parameters according to a set time frequency, and arrange the sampled environmental parameters according to the sampling time sequence to form meteorological time series data; Interpolate missing points in the meteorological time series data; Remove outliers in the meteorological time series data; Normalize the meteorological time series data; Perform wavelet transform decomposition on the normalized meteorological time series data to obtain training data of the meteorological time series data. 3.The short-term wind-solar power output prediction method based on the dual-branch deep learning model according to claim 1, characterized in that: The preprocessing of the meteorological spatial image data comprises: The meteorological spatial image data is a satellite cloud image or a weather radar echo image; Align the meteorological spatial image data and the meteorological time series data by time stamp to ensure their consistency in the time scale; Crop the meteorological spatial image data; Scale the size of the cropped meteorological spatial image data and unify it to a set size; Normalize the pixel values of the size-scaled meteorological spatial image data; Constructing an image sequence tensor: ; wherein: T is the number of time steps, H, W and C are the height, width and channel number of the image respectively; the image sequence tensor is used as pre-processed meteorological space image data for training of the dual-branch deep learning model. 4.The short-term wind-solar power output prediction method based on the dual-branch deep learning model according to claim 2, characterized in that: The prediction output network comprises a regression prediction network composed of fully connected layers and a wavelet inverse transform module; The regression prediction network outputs wavelet component features corresponding to the wind and light output, and the wavelet inverse transform module inversely transforms the wavelet component features to obtain the final wind and light output result. 5.The short-term wind-solar power output prediction method based on the dual-branch deep learning model according to claim 1, characterized in that: The calculation process of the LSTM time feature extraction network is: forget gate computation: Input gate computation: Candidate cell state computation: Cell status update: Output gate computation: Hidden state output: ; where: is the input vector at the current time step, is the hidden state at the previous time step, is a sigmoid activation function that compresses the gating signal to the interval (0, 1), and denotes element-wise multiplication.