A short-term photovoltaic power prediction method and system based on a mixture model

Abstract

The application provides a short-term photovoltaic power prediction method and system based on a hybrid model, obtains data related to photovoltaic power, constructs a hybrid model of a three-dimensional convolution and a convolution long short-term memory neural network, trains the hybrid model, and obtains a short-term photovoltaic power prediction result based on the data related to photovoltaic power by using the trained hybrid model. The prediction result of the application has higher accuracy and shorter time consumption, and proves the applicability, accuracy and stability of the constructed model for photovoltaic power prediction.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power prediction technology, specifically relating to a short-term photovoltaic power prediction method and system based on a hybrid model. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Photovoltaic power generation is characterized by randomness and intermittency. Its output is significantly affected by factors such as solar irradiance, temperature, humidity, shading, haze particles, the aging of photovoltaic panels, and load conditions. This poses a challenge to the accurate prediction of photovoltaic power generation. However, accurate photovoltaic power generation prediction is crucial for ensuring the safe and stable operation of the power system and for effective coordinated dispatching. Therefore, further improving the accuracy of photovoltaic power prediction is of paramount importance. Photovoltaic power prediction can be categorized into three main types: physical methods based on physical data modeling, statistical methods based on numerical analysis, and machine learning.

[0004] The physical method establishes a physical model based on meteorological factors such as solar irradiance, temperature, humidity, and wind speed, combined with indirect influencing factors such as the installation angle of the photovoltaic panel, photoelectric conversion efficiency, and battery conditions. The physical prediction model does not require historical data to predict the photovoltaic output power. The prediction accuracy depends on a large amount of meteorological data, detailed geographical location information, and complete technical parameters. However, meteorological data is not easy to obtain, and the accuracy and quantity of parameters provided by photovoltaic panel manufacturers are limited. Furthermore, the established model is limited by regional empirical parameters, resulting in poor anti-interference ability and universality of the model, and weak robustness.

[0005] Statistical methods rely on numerical analysis of historical data such as solar irradiance and photovoltaic power. This includes linear fitting of historical data curves, numerical calculation of relevant parameters, and correlation analysis among various influencing factors. These methods establish a correspondence between input and output data (i.e., a data model) to predict future photovoltaic output power. Commonly used methods include time series analysis, fuzzy theory, grey theory, spatiotemporal correlation analysis, and regression analysis. However, statistical methods require large amounts of historical data and are time-consuming. In short-term predictions, the confidence level of the values ​​deteriorates. Furthermore, the prediction quality of statistical methods is closely related to the quality of the data and the availability of historical data. The effectiveness of data screening and the elimination of false data significantly impacts prediction accuracy, making it difficult to guarantee accuracy.

[0006] Machine learning utilizes activation functions to uncover complex nonlinear characteristics among relevant factors, overcoming the limitations of linear models in representation and classification. Prediction methods based on various neural networks have become one of the most commonly used mainstream methods for photovoltaic power prediction in recent years, and are widely applied in power system power forecasting.

[0007] According to the inventors, some literature proposes a photovoltaic power prediction model based on artificial neural networks, using meteorological data (such as solar irradiance, wind speed, humidity, temperature, and air pressure) as input data. The model is compared with traditional physical models and typical statistical models, showing that the selected model yields the best prediction results, but the prediction process is time-consuming. Other literature uses support vector machines to predict photovoltaic power, demonstrating that changes in meteorological factors significantly impact photovoltaic power output, but the algorithm is complex. Still other literature uses irradiance, temperature, humidity, wind speed, and air pressure as input data to validate single convolutional neural network (CNN), long short-term memory (LSTM), and hybrid models combining both for photovoltaic power prediction. The results show that the hybrid model has better prediction accuracy than either of the two single models. Some literature indicates that deep neural networks have better feature extraction capabilities than shallow neural networks, improving the gradient vanishing problem during neural network operation, thus exhibiting better performance and robustness. However, it does not fully explore the impact of complex nonlinear characteristics of meteorological data on photovoltaic output power. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a short-term photovoltaic power prediction method and system based on a hybrid model. The prediction results of this invention are not only more accurate but also less time-consuming, demonstrating the applicability, accuracy, and stability of the established model for photovoltaic power prediction.

[0009] According to some embodiments, the present invention adopts the following technical solution:

[0010] A short-term photovoltaic power prediction method based on a hybrid model includes the following steps:

[0011] Obtain data related to photovoltaic power;

[0012] Construct a hybrid model of 3D convolutional and convolutional long short-term memory neural networks;

[0013] Train the hybrid model;

[0014] Using the trained hybrid model, short-term photovoltaic power prediction results are obtained based on data related to photovoltaic power.

[0015] As an alternative implementation method, data related to photovoltaic power include solar irradiance, temperature, humidity, and wind speed.

[0016] As an alternative implementation, the acquired data may also be preprocessed, the preprocessing process including deleting abnormal data and filling in missing data.

[0017] As an optional implementation, the preprocessing process further includes forming all the data into a vector, multiplying it by its transpose to obtain a two-dimensional matrix, acquiring data once every set time interval, and merging the collected data to form a three-dimensional matrix as the input vector of the model.

[0018] As an alternative implementation, the hybrid model includes an improved three-dimensional convolutional neural network and a convolutional long short-term memory neural network. The improved three-dimensional convolutional neural network uses a standard convolutional layer to replace the pooling layer and two consecutive convolutional layers to form the three-dimensional convolutional neural network, thereby refining the characteristics of photovoltaic power-related factors.

[0019] As a further implementation, the improved three-dimensional convolutional neural network performs at least two consecutive convolution processes, using a three-dimensional convolutional kernel to perform convolution operations on the input matrix, and using the output of the first convolutional layer as input to obtain the second convolutional layer.

[0020] Each convolutional layer is followed by a batch normalization layer and a non-linear activation function layer, and the weight coefficients of the three-dimensional convolutional kernels are the same;

[0021] Furthermore, a padding operation is performed before the convolution operation.

[0022] As a further implementation, the excitation function of the excitation layer is ReLU.

[0023] As a further implementation, the convolutional long short-term memory neural network includes an input gate, a forget gate, a cell, an output gate, and a hiding place, and uses convolution operations to connect the transformations between different gates.

[0024] As an alternative implementation, the training process of the hybrid model includes:

[0025] The batch training method was adopted, and multiple meteorological data were randomly collected as the training set.

[0026] Random initialization is performed, setting the shared weight range of the hybrid network to values ​​close to zero in [0, 1], and initializing the learning rate η (η = 0.1) and training error threshold ε;

[0027] Calculate the error loss of the training set based on the selected loss function;

[0028] The error values ​​are backpropagated using batch stochastic gradient descent to update the network parameters.

[0029] Determine if the total error E of the model after weight adjustment is less than ε. If E < ε, proceed to the next step; return to continue training.

[0030] After training, the network parameters are saved to obtain the optimized network model.

[0031] A short-term photovoltaic power prediction system based on a hybrid model includes:

[0032] The data acquisition module is configured to acquire data related to photovoltaic power.

[0033] The model building module is configured to build a hybrid model of 3D convolutional and convolutional long short-term memory neural networks;

[0034] The model training module is configured to train the hybrid model;

[0035] The power prediction module is configured to use a trained hybrid model to obtain short-term photovoltaic power prediction results based on data related to photovoltaic power.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] By incorporating a convolutional neural network, this invention can refine the interaction relationships between various parameters. In addition to extracting features in the time dimension, it can also extract spatial features of the data and is more sensitive to changes in influencing factors, making the model's prediction results more accurate.

[0038] To make the above-mentioned objects, features and advantages of the present invention clearer and easier to understand, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0040] Figure 1 This is a diagram of the 3DCNN+CLSTM hybrid model structure in this embodiment;

[0041] Figure 2 This is a schematic diagram of the convolution process in this embodiment;

[0042] Figure 3 This is a schematic diagram of the internal structure of the CLSTM in this embodiment;

[0043] Figure 4 This is a schematic diagram of the measured photovoltaic power and solar irradiance in this embodiment;

[0044] Figure 5 This is a schematic diagram of the method flow in this embodiment;

[0045] Figure 6 This is a schematic diagram of the model convergence process in this embodiment;

[0046] Figures 7(a) and 7(b) show a comparison of the photovoltaic power generation prediction results for sunny and sudden weather conditions in this embodiment, respectively.

[0047] Figure 8 The RMES, MEA, and MAPE values ​​for the clear-sky power prediction results in this embodiment are:

[0048] Figure 9 This example compares the evaluation indicators under two different weather conditions. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0052] Example 1

[0053] A photovoltaic power prediction model based on 3D convolutional neural networks (3DCNN) and convolutional long short-term memory neural networks (CLSTM) can fully exploit the complex nonlinear relationships between input features. The proposed model is compared with single traditional backpropagation neural network and convolutional long short-term memory neural network models, as well as 3D convolutional neural network models. The model shows higher prediction accuracy and shorter prediction time, demonstrating the applicability, accuracy, and stability of the proposed model for photovoltaic power prediction.

[0054] The model structure proposed in this embodiment is as follows: Figure 1 As shown, the model consists of a three-dimensional convolutional neural network and a convolutional long short-term memory neural network. In order to gain a deeper understanding of the internal logical structure of the model, a detailed description is provided below.

[0055] 3D Convolutional Neural Network

[0056] The traditional construction principle of 3D convolutional neural networks involves alternating between convolutional and pooling layers, followed by a small number of fully connected layers. Convolutional neural networks used for object and image recognition all include pooling layers. The main function of pooling layers is feature dimensionality reduction, compressing the amount of computational data and related parameters, reducing model computational complexity, decreasing model runtime, and improving the model's fault tolerance during computation. Therefore, the presence of pooling layers cannot meet the requirement of fully exploring the nonlinear relationships between factors influencing photovoltaic power generation in photovoltaic power prediction.

[0057] This embodiment selects meteorological data used in the experiment that exhibit complex nonlinear relationships. The established three-dimensional convolutional model replaces the pooling layer with a standard convolutional layer, and uses two consecutive convolutional layers to construct a three-dimensional convolutional neural network, refining the characteristics of photovoltaic power-related factors. Removing the pooling layer does not affect the stability and accuracy of the convolutional neural network. Replacing the pooling layer with a standard convolutional layer fully explores the nonlinear relationships of various influencing factors, making the model more practical.

[0058] Data processing

[0059] Before data is transmitted to the network, it needs to be preprocessed, including removing outlier data (in extreme cases) and filling in missing data (such as equipment malfunctions). Training deep learning networks requires a large amount of experimental data, so data augmentation is employed. Data normalization is used to unify the data to the same units of measurement. Because 3DCNN+CLSTM has specific requirements for the format of input data, it is necessary to convert the data to an appropriate format.

[0060] This embodiment selects solar irradiance, temperature, humidity, and wind speed as input data. A key feature distinguishing CNNs from other models is that their input data is typically required to be a three-dimensional matrix. The output and input data are directly connected through a convolution operation. Features are obtained during convolution by sharing convolution kernels (weights and biases). First, the original vector composed of the input data is multiplied by its transpose to obtain a 4×4 two-dimensional matrix. Data is collected at regular intervals (e.g., 5 minutes), and the data collected at each set interval (e.g., 60 minutes) is combined into a single three-dimensional matrix as the model's input vector. Assume z = (a, b, c, d). T ,but:

[0061]

[0062] In the formula: a represents solar irradiance, b represents temperature, c represents wind speed, and d represents humidity.

[0063] Convolutional layer

[0064] Convolutional layers are the core of convolutional neural networks. Convolution is a local operation; it uses a kernel of a certain size to perform a convolution operation on the input vector to obtain local information about the input vector. Local connectivity and weight sharing are two important characteristics. Two consecutive convolution processes in the model are as follows: Figure 2 As shown.

[0065] The specific process of the two convolution operations is as follows: After forming a three-dimensional input vector, it is then subjected to convolution operations on the input matrix by four 1×2×2 (depth, length, width) three-dimensional convolution kernels. "1" represents the depth of the three-dimensional convolution kernel (the time dimension of the three-dimensional convolution kernel), and "1" represents the number of input matrices processed by the model each time. Four 2×2 spatial windows slide sequentially from top to bottom and from left to right. The convolution operation of the four three-dimensional convolution kernels yields four 10×3×3 three-dimensional matrices, which are the output data of the first convolutional layer (Conv3D_1). Conv3D_1 has 20 parameters. The second convolution operation is similar to the first, using the output of the first convolutional layer as input to obtain the second convolutional layer (Conv3D_2). Conv3D_2 has 68 parameters. The expressions for the two convolution operations are:

[0066]

[0067] In the formula: ReLU is the activation function of the model, expressed as f(x) = max(0,x), Z ij Represents the input vector, X xyz This is the calculation result of the first convolutional layer, where x and y are the position coordinates of the output data in each layer, and z represents the channel depth (layer). It is the shared weight of each layer, and its value range is (0,1), where l m If the output matrix channel depth is 'b', then each 2×2 convolutional kernel has 4 shared weights with respect to the input vector. lm This represents the shared bias within the range (0,1).

[0068] Each Conv3D layer is followed by a batch normalization (BN) layer and a non-linear activation function layer. Each 3D convolutional kernel has the same weight coefficients. To keep the feature map size constant, padding is performed before the convolution operation.

[0069] The main purpose of the activation layer in a three-dimensional convolutional neural network is to improve the nonlinear expressive power of the neural network. The activation function selected in this paper is ReLU. The characteristic of this function is that there is no saturation region during the model operation. It also solves the gradient vanishing problem and prevents gradient diffusion. It does not have a complicated operation process and its efficiency is greatly improved. Compared with the sigmoid activation function and tanh activation function, it has a faster convergence speed.

[0070] The process of photovoltaic power prediction is a problem of processing spatiotemporal series data. The established model performs two three-dimensional convolution operations on the input data, and finally obtains 12 2×2×4 three-dimensional matrices arranged in time series. The output data obtained from the two convolutions is unfolded according to the time dimension and used as the input data of the convolutional long short memory layer, finally obtaining 4 two-dimensional matrices. The 4 two-dimensional matrices are stretched into a vector and sent to the fully connected layer for operation.

[0071] LSTM networks consist of five modules: input gate, forget gate, unit, output gate, and hidden layer. LSTM typically uses fully connected layers to connect the transitions between different gates, while CLSTM uses convolutional operations to connect the transitions between different gates. The internal structure of a convolutional long short-term memory neural network is as follows: Figure 3 As shown.

[0072] Formula (3) illustrates the working principle of CLSTM, as shown below:

[0073]

[0074] In the formula: i t f t g t C t O t H t These correspond to six modules: input gate, forget gate, input modulation gate, unit, output gate, and hiding gate. " indicates matrix multiplication; "*" indicates convolution; "σ" and "tanh" are activation functions; W xi These represent the weights of the modules mentioned above; b i For shared bias; t represents the time series.

[0075] Simulation Examples

[0076] To evaluate the predictive performance of the model proposed in the above embodiments, three accuracy evaluation metrics were selected in the prediction experiments. These are the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), specifically:

[0077]

[0078] Loss is the loss function chosen in this embodiment.

[0079]

[0080] In the formula, y(i) is the actual photovoltaic output power data, y'(i) is the predicted value, and N is the sequence number of the measured data y(i).

[0081] Data from a photovoltaic power station in Jingmen, Hubei Province, was selected for a case study analysis. Specific information about the system is shown in Table 1, which displays detailed information about the solar energy device. This experiment used data from a certain month in 2021 as the training set, with sampling time from 5:00 AM to 7:00 PM daily and a data resolution of 5 minutes. Two days (sunny and with sudden weather changes) were selected as the test set. The data downloaded from the research institute mainly included active power (kW), wind speed (m / s), temperature (°C), relative humidity (%), and solar irradiance (kW / m²). 2 ).

[0082] Table 1 System Details

[0083]

[0084] Figure 4 The graph displays historical data on measured photovoltaic power generation and solar irradiance for a given week. The horizontal axis represents time, the left vertical axis represents power, and the right vertical axis represents solar irradiance. As can be seen from the graph, photovoltaic output varies with irradiance. Power and solar irradiance are highest at noon, relatively low in the morning and afternoon, and approach zero at night.

[0085] In the simulation experiments, a total of four simulations were conducted to verify the performance of the proposed model and the impact of the input sequence on the prediction accuracy of the hybrid model. These simulations included the BP neural network model, the CLSTM model, the 3DCNN model, and the Conv3D+CLSTM hybrid model. To verify the effectiveness of the model, the four models were input with data of the same sequence length during the experiment, and the training and optimization methods were also consistent with the hybrid model. The training and validation cycle of the model was 100, and the model code was built in Python 3.8.

[0086] The training process of the 3DCNN+CLSTM hybrid model is as follows:

[0087] (1) This experiment uses the batch training method, randomly collecting 5000 meteorological data points on a certain day as the training set;

[0088] (2) Random initialization: Set the shared weight range of the hybrid network to a value close to zero in [0, 1], and initialize the learning rate η (η = 0.1) and the training error threshold ε;

[0089] (3) Calculate the error loss of the training set based on the selected loss function;

[0090] (4) Backpropagate the error values ​​using the batch stochastic gradient descent method to update the network parameters;

[0091] (5) Determine whether the total error E of the model after adjusting the weights is less than ε. If E < ε, proceed to the next step; otherwise, return to step 3 and continue training.

[0092] (6) After training, save the network parameters to obtain the optimized network model. Specific information is as follows: Figure 5 As shown.

[0093] After 100 iterations, the error thresholds of the four models met the initial settings. The Adam optimization algorithm was used to verify the convergence of all four models. The convergence process of the four models is detailed below. Figure 6 . Figure 6 The horizontal axis represents the number of model iterations, and the vertical axis represents the MSE values ​​of the four models.

[0094] from Figure 6 It can be seen that the BP neural network model has greater volatility in the early stage, and the four models have basically the same trend. However, it is clear that the convergence of the 3DCNN+CLSTM model is better than the other three models, and the convergence speed of the 3DCNN+CLSTM model is the fastest.

[0095] Figures 7(a) and 7(b) show the photovoltaic power generation prediction results for two different weather conditions under the same input sequence after the four models have been trained. Figure 7(a) represents a sunny day, and Figure 7(b) represents a sudden weather change. As can be seen from the figures, regardless of whether it is a sunny day or a sudden weather change with meteorological fluctuations, all four models can reflect the trend of photovoltaic power changes within a day. However, the prediction results of the BP neural network model show significant fluctuations compared to the measured data, indicating low prediction accuracy. This shows that the BP model cannot accurately reflect the changes in photovoltaic power in a timely manner during the prediction process. Backpropagation (BP) neural networks have highly complex nonlinear structures, which can easily lead to local minima during training and learning, preventing them from converging quickly to reach the global optimum. This results in the final output value not meeting expectations. Individual CLSTM and 3DCNN models outperform BP neural network models in prediction, but they are slightly inferior to the 3DCNN+CLSTM model. This is mainly because the addition of convolutional neural networks can refine the interaction between parameters, extract spatial features in addition to temporal features, and is more sensitive to changes in influencing factors, making the model's prediction results more accurate.

[0096] Figure 8 The table shows a comparison of the RMSE, MAE, and MAPE results of the power prediction results of the four models under clear weather conditions. As can be seen from the table, the model proposed in this paper outperforms the other three single models in terms of prediction results under clear weather conditions.

[0097] Figure 9The graph compares the MAE, MAPE, and RMSE of different models under two different weather conditions. Uppercase letters represent sudden weather changes, and lowercase letters represent sunny weather. As can be seen from the graph, the prediction accuracy of the 3DCNN+CLSTM model is better than that of the selected comparison model under both weather conditions. Therefore, it can be concluded that the deep learning prediction model has strong robustness and good performance in photovoltaic power prediction. The same performance evaluation index of the same model under the two weather conditions has slight errors, which is due to the change of hyperparameters in the training results, and is within a controllable range.

[0098] Overall, the 3DCNN+CLSTM hybrid model outperforms the three individual models in prediction. This is because the CLSTM model cannot fully exploit the nonlinear relationships between input vectors, resulting in poorer predictions compared to the 3DCNN+CLSTM model. The 3DCNN model, being a spatially deep neural network, primarily extracts the spatial depth features of the input vectors, but it cannot fully capture the interactions between various influencing factors and data fluctuations. Therefore, the hybrid model's prediction results are superior to those of a single 3DCNN model. The comparison shows that the 3DCNN+CLSTM model outperforms the other three models. This high stability and robustness is because the model refines the complex nonlinear relationships between influencing factors and photovoltaic power generation, extracting the spatiotemporal features of the data, ultimately achieving higher prediction accuracy.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A short-term photovoltaic power prediction method based on a hybrid model, characterized in that, Includes the following steps: Obtain data related to photovoltaic power; Construct a hybrid model of 3D convolutional and convolutional long short-term memory neural networks; Train the hybrid model; Using the trained hybrid model, short-term photovoltaic power prediction results are obtained based on data related to photovoltaic power. The hybrid model includes an improved three-dimensional convolutional neural network and a convolutional long short-term memory neural network. The improved three-dimensional convolutional neural network uses a standard convolutional layer to replace the pooling layer and uses two consecutive convolutional layers to form a three-dimensional convolutional neural network, which refines the characteristics of photovoltaic power-related factors. The improved three-dimensional convolutional neural network performs at least two consecutive convolution processes, using a three-dimensional convolutional kernel to perform convolution operations on the input matrix, and using the output of the first convolutional layer as input to obtain the second convolutional layer; Each convolutional layer is followed by a batch normalization layer and a non-linear activation function layer, and the weight coefficients of the three-dimensional convolutional kernels are the same; Perform padding before the convolution operation; The model performs two three-dimensional convolution operations on the input data, resulting in 12 2×2×4 three-dimensional matrices arranged in time sequence. The output data obtained from the two convolutions is unfolded along the time dimension and used as input data for the convolutional long short memory layer, resulting in four two-dimensional matrices. These four two-dimensional matrices are stretched into a vector and sent to the fully connected layer for computation.

2. The short-term photovoltaic power prediction method based on a hybrid model as described in claim 1, characterized in that, Data related to photovoltaic power includes solar irradiance, temperature, humidity, and wind speed.

3. The short-term photovoltaic power prediction method based on a hybrid model as described in claim 1, characterized in that, It also includes preprocessing the acquired data, the preprocessing process including deleting abnormal data and filling in missing data.

4. A short-term photovoltaic power prediction method based on a hybrid model as described in claim 1 or 3, characterized in that, The preprocessing process also includes forming all the data into a vector, multiplying it by its transpose to obtain a two-dimensional matrix, acquiring data once every set time interval, and merging the collected data to form a three-dimensional matrix as the input vector of the model.

5. The short-term photovoltaic power prediction method based on a hybrid model as described in claim 1, characterized in that, The excitation function of the excitation layer is ReLU.

6. The short-term photovoltaic power prediction method based on a hybrid model as described in claim 1, characterized in that, The convolutional long short-term memory neural network includes an input gate, a forget gate, a unit, an output gate, and a hidden gate, and uses convolution operations to connect the transformations between different gates.

7. The short-term photovoltaic power prediction method based on a hybrid model as described in claim 1, characterized in that, The training process of the hybrid model includes: The batch training method was adopted, and multiple meteorological data were randomly collected as the training set. Random initialization is performed, setting the shared weights of the hybrid network to values ​​close to zero in the range [0, 1], and initializing the learning rate. η ( η =0.1) and training error threshold ε ; Calculate the error loss of the training set based on the selected loss function; The error values ​​are backpropagated using batch stochastic gradient descent to update the network parameters. Determine whether the total error E of the model after adjusting the weights is less than ε If E < ε If yes, proceed to the next step; return to continue training. After training, the network parameters are saved to obtain the optimized network model.

8. A short-term photovoltaic power prediction system based on a hybrid model, characterized in that, include: The data acquisition module is configured to acquire data related to photovoltaic power. The model building module is configured to build a hybrid model of 3D convolutional and convolutional long short-term memory neural networks; The model training module is configured to train the hybrid model; The power prediction module is configured to use a trained hybrid model to obtain short-term photovoltaic power prediction results based on data related to photovoltaic power. The hybrid model includes an improved three-dimensional convolutional neural network and a convolutional long short-term memory neural network. The improved three-dimensional convolutional neural network uses a standard convolutional layer to replace the pooling layer and uses two consecutive convolutional layers to form a three-dimensional convolutional neural network, which refines the characteristics of photovoltaic power-related factors. The improved three-dimensional convolutional neural network performs at least two consecutive convolution processes, using a three-dimensional convolutional kernel to perform convolution operations on the input matrix, and using the output of the first convolutional layer as input to obtain the second convolutional layer; Each convolutional layer is followed by a batch normalization layer and a non-linear activation function layer, and the weight coefficients of the three-dimensional convolutional kernels are the same; Perform padding before the convolution operation; The model performs two three-dimensional convolution operations on the input data, resulting in 12 2×2×4 three-dimensional matrices arranged in time sequence. The output data obtained from the two convolutions is unfolded along the time dimension and used as input data for the convolutional long short memory layer, resulting in four two-dimensional matrices. These four two-dimensional matrices are stretched into a vector and sent to the fully connected layer for computation.

Images