Short-term photovoltaic power generation power time sequence prediction method and system for household energy

Abstract

The invention belongs to the technical field of power generation power prediction, and discloses a short-term photovoltaic power generation power time sequence prediction method and system for household energy, and the method comprises the steps: obtaining an image sequence matched with photovoltaic power generation data from foundation cloud picture data based on a time data input sequence of the preprocessed photovoltaic power generation data; splicing the image sequence after dimension reduction with photovoltaic power generation data to form a time feature sequence; decomposing the time feature sequence to obtain a trend sequence and a residual sequence; and in the DLinear structure, constructing a full connection relationship among the trend sequence, the residual sequence and the to-be-predicted generated power sequence, and enabling the full connection relationship to directly output the to-be-predicted generated power sequence through the plurality of decomposed sequences. According to the method, the foundation cloud picture and the photovoltaic data are fused, the spatial-temporal features are extracted in combination with the 3D convolutional network, the time sequence trend and fluctuation are decomposed by using DLinear, and the photovoltaic power generation power prediction precision is remarkably improved through global association and adaptive weight adjustment.

Description

Technical Field

[0001] The invention belongs to the technical field of power generation prediction, and relates to a method and system for predicting short-term photovoltaic power generation time series of household energy. Background Art

[0002] With the growth of global energy demand and the intensification of climate change issues, renewable energy, especially solar photovoltaic power generation, has received widespread attention. Under the current trend of full electrification of distributed energy, the energy system is gradually integrating diversified flexible resources. The scheduling of flexible resources in the future can improve the energy utilization of the system to a certain extent. As the main source of flexible resources, photovoltaic systems are subject to large fluctuations in a short period of time due to the movement of cloud clusters. Accurately predicting their short-term power generation is critical to establishing the charging and discharging control strategy of energy storage systems.

[0003] The current short-term photovoltaic power prediction method mainly adopts the machine learning method. At the input end of the model, physical parameters or cloud map data are usually selected as the model input. Some scholars have proposed methods for short-term photovoltaic power prediction, but the prediction premise of such methods is to build a complex data link. At the same time, for household energy systems, their distribution locations are not fixed. If the meteorological station data in the area is selected, the error is large and it is impossible to accurately predict the future photovoltaic volatility.

[0004] In the prediction method based on cloud image data, the data set usually uses ground-based cloud image data obtained by an all-sky imager in conjunction with a complex shading system as image input. This method has a large initial investment and is often used for site-level photovoltaic prediction. It is not suitable for household energy systems, which are smaller in scale and have limited initial equipment investment. Summary of the invention

[0005] The purpose of the present invention is to solve the problem in the prior art that short-term photovoltaic power generation prediction requires the construction of complex data links, has large initial investment, and is not suitable for household energy systems, and to provide a method and system for timing prediction of short-term photovoltaic power generation for household energy.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A short-term photovoltaic power generation time series prediction method for household energy, comprising:

[0008] Collect ground-based cloud map data and photovoltaic power generation data;

[0009] Preprocessing the collected photovoltaic power generation data, and obtaining an image sequence matching the photovoltaic power generation data in the ground-based cloud image data based on the time data input sequence of the preprocessed photovoltaic power generation data;

[0010] Reduce the dimension of the image sequence based on a 3D convolutional neural network, and splice the reduced image sequence with photovoltaic power generation data to form a time feature sequence;

[0011] Input the time feature sequence into the DLinear structure for decomposition to obtain a trend sequence and a residual sequence, and in the DLinear structure, construct a fully connected relationship between the trend sequence, the residual sequence and the power generation power sequence to be predicted, so that it directly outputs the power generation power sequence to be predicted through the decomposed multiple sequences.

[0012] A further improvement of the present invention lies in:

[0013] Furthermore, preprocess the collected photovoltaic power generation data, specifically:

[0014] Detect and remove outliers in the photovoltaic power generation data based on the 3σ rule, and fill in missing values by linear interpolation;

[0015] The ground-based cloud map data is obtained by taking pictures of the sky cloud map with a household fisheye camera.

[0016] Furthermore, based on the time data input sequence of the preprocessed photovoltaic power generation data, obtain an image sequence matching the photovoltaic power generation data in the ground-based cloud map data, specifically:

[0017] Determine the time range of the image sequence to be matched according to the time stamp of the time data input sequence;

[0018] Extract key frames from the images in the ground-based cloud map dataset at fixed frame intervals and save them in image format;

[0019] Sort the extracted images according to the time stamp, and filter out the image sequence corresponding to the time range of the time data input sequence;

[0020] Transform the filtered image sequence based on the nearest neighbor interpolation method to obtain images with a unified size.

[0021] Furthermore, the time data input sequence is specifically:

[0022] Intercept a fixed time period, obtain the data volume contained in the fixed time period, and filter out the data volume corresponding to the window size from the fixed time period through a sliding window;

[0023] Based on the set prediction time scale, construct a corresponding time data input sequence;

[0024] Among them, constructing the corresponding data input sequence is: taking the data points corresponding to the prediction time scale as the quantity to be predicted, and taking the data points corresponding to the difference between the sliding window size and the prediction time scale as the time data input sequence.

[0025] Further, dimensionality reduction is performed on the image sequence based on a 3D convolutional neural network, specifically: reducing the image sequence data with the specification of L*C*H*W to the specification of L*N, where L is the length of the image sequence, C is the number of image channels, H is the image height, W is the image width, and N is the number of extracted features.

[0026] 10. The short-term photovoltaic power time series prediction method for household energy according to claim 5, characterized in that the 3D convolutional neural network includes four consecutive 3D convolutional layers, and a ReLU activation function and a Dropout layer are connected after each 3D convolutional layer; the 3D convolutional layers include: Conv1, Conv2, Conv3, and Conv4; the 3D convolutional layers reduce the image sequence to a feature sequence of 8 parameters;

[0027] Conv1 has an input channel number of 3, an output channel number of 32, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0028] Conv2 has an input channel number of 32, an output channel number of 16, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0029] Conv3 has an input channel number of 16, an output channel number of 8, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0030] Conv4 has an input channel number of 8, an output channel number of 8, a convolutional kernel size of (3, 2, 2), a stride of (1, 2, 2), and a padding of (1, 0, 0).

[0031] Further, the time feature sequence is input into the DLinear structure for decomposition to obtain a trend sequence and a residual sequence, specifically: the time feature sequence is input into the DLinear structure, and each variable time series is decomposed into a trend sequence and a residual sequence. The sequence part calculated by moving average represents the relatively smooth part in the data and is used as the long-term trend of the time series; the residual sequence is the rapidly changing component and noise.

[0032] Further, in the DLinear structure, a fully connected relationship among the trend sequence, the residual sequence, and the power generation power sequence to be predicted is constructed, enabling it to directly output the power generation power sequence to be predicted through the decomposed multiple sequences. Specifically: the time feature sequence is a time sequence with 9 original feature parameters; when the time feature sequence is decomposed into a trend sequence and a residual sequence, each original time sequence will generate two new sequences, a trend sequence and a residual sequence; for the 9 original time sequences, 2 * 9 = 18 new sequences will be obtained; a fully connected relationship is established between the 18 newly generated sequences and the power generation power sequence to be finally predicted; the established fully connected relationship connects the 18 * L parameters of the 18 new sequences with the respective parameters on the power generation power sequence to be predicted according to the randomly assigned initial weights; that is, the parameter points at each moment of each new sequence are connected to the parameter points at each moment of the power generation power sequence to be predicted.

[0033] The 18 * L parameter points of the new sequences are mapped to the L parameter points of the power generation power sequence to be predicted through the fully connected layer; the parameter points at each moment of the power generation power sequence to be predicted are the weighted sum of the parameter points at all moments of the new sequences; calculate the mean square error MSE between the power generation power sequence to be predicted and the historical power generation power sequence; calculate the gradient of the loss function with respect to the weights through the backpropagation algorithm; use the optimizer to update the weights according to the gradient; repeatedly perform forward propagation, loss calculation, backpropagation, and weight update until the model converges or reaches the predetermined number of training epochs; and adaptively adjust the weight distribution during iterative training. The trained model directly outputs the predicted photovoltaic power generation sequence.

[0034] A short-term photovoltaic power generation power time series prediction system for household energy, comprising:

[0035] An acquisition module, which acquires ground-based cloud map data and photovoltaic power generation data;

[0036] A preprocessing module, which preprocesses the acquired photovoltaic power generation data, and based on the time data input sequence of the preprocessed photovoltaic power generation data, obtains an image sequence matching the photovoltaic power generation data in the ground-based cloud map data;

[0037] A dimensionality reduction module, which reduces the dimension of the image sequence based on a 3D convolutional neural network and stitches the dimension-reduced image sequence with the photovoltaic power generation data to form a time feature sequence;

[0038] A decomposition module, which inputs the time feature sequence into the DLinear structure for decomposition to obtain a trend sequence and a residual sequence, and in the DLinear structure, constructs a fully connected relationship among the trend sequence, the residual sequence, and the power generation power sequence to be predicted, enabling it to directly output the power generation power sequence to be predicted through the decomposed multiple sequences.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] Based on the time data input sequence, the present invention accurately matches the image sequence in the ground cloud map data, realizing the accurate association between photovoltaic power generation and weather conditions; and through the dimensionality reduction processing of the 3D convolutional neural network, not only significantly reduces the data dimension, but also retains the key features, significantly improving the calculation efficiency. The dimensionality-reduced image sequence is spliced with the photovoltaic power generation data to form a time feature sequence, which comprehensively reflects the spatio-temporal characteristics of photovoltaic power generation. The decomposition strategy of the DLinear structure effectively separates the trend sequence and the residual sequence, enabling the prediction model to more accurately capture the long-term trend and short-term fluctuations of photovoltaic power generation. The prediction model established by relying on the fully connected relationship inside can accurately output the predicted power generation power sequence, effectively improving the operation efficiency and economic benefits of the photovoltaic power generation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0042] Figure 1 It is a schematic flow chart of a method for predicting the short-term photovoltaic power generation power time series of a household energy of the present invention;

[0043] Figure 2 It is a schematic structural diagram of a system for predicting the short-term photovoltaic power generation power time series of a household energy of the present invention;

[0044] Figure 3 It is a schematic diagram of the dimensionality reduction of the image sequence by the 3D convolutional neural network;

[0045] Figure 4 It is a schematic structural diagram of the 3D convolutional neural network;

[0046] Figure 5 It is a schematic flow chart of the output process of the photovoltaic power generation power sequence to be predicted. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0048] Accordingly, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts fall within the scope of protection of the present invention.

[0049] It should be noted that like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not require further definition and explanation in subsequent drawings.

[0050] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper", "lower", "horizontal", "inner", etc. are used to indicate the orientation or positional relationship, it is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the invention is customarily placed. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention. In addition, terms such as "first", "second", etc. are only used for descriptive distinction and cannot be construed as indicating or implying relative importance.

[0051] In addition, if the term "horizontal" appears, it does not mean that the component is required to be absolutely horizontal, but it can be slightly inclined. For example, "horizontal" only means that its direction is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but it can be slightly inclined.

[0052] In the description of the embodiments of the present invention, it should also be noted that unless otherwise clearly specified and limited, if terms such as "set", "installed", "connected", "connected" are used, they should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0053] The present invention will be further described in detail below with reference to the accompanying drawings:

[0054] See Figure 1 , the present invention discloses a short-term photovoltaic power time series prediction method for household energy, including:

[0055] S101, collecting ground-based cloud map data and photovoltaic power generation data;

[0056] S102. Preprocess the collected photovoltaic power generation data. Based on the time data input sequence of the preprocessed photovoltaic power generation data, obtain an image sequence that matches the photovoltaic power generation data from the ground-based cloud map data;

[0057] Preprocess the collected photovoltaic power generation data specifically as follows:

[0058] Detect and remove outliers in the photovoltaic power generation data based on the 3σ rule, and fill in missing values by linear interpolation;

[0059] The ground-based cloud map data is obtained by taking pictures of the sky cloud map with a household fisheye camera.

[0060] Based on the time data input sequence of the preprocessed photovoltaic power generation data, obtain an image sequence that matches the photovoltaic power generation data from the ground-based cloud map data specifically as follows:

[0061] Determine the time range of the image sequence to be matched according to the timestamps of the time data input sequence;

[0062] Extract key frames from the images in the ground-based cloud map dataset at fixed frame intervals and save them in image format;

[0063] Sort the extracted images according to the timestamps and filter out the image sequence corresponding to the time range of the time data input sequence;

[0064] Transform the filtered image sequence based on the nearest neighbor interpolation method to obtain images of uniform size.

[0065] The time data input sequence is specifically as follows:

[0066] Intercept a fixed time period, obtain the data volume contained in the fixed time period, and filter out the data volume corresponding to the window size from the fixed time period through a sliding window;

[0067] Based on the set prediction time scale, construct a corresponding time data input sequence;

[0068] Among them, constructing the corresponding data input sequence is: taking the data points corresponding to the prediction time scale as the quantity to be predicted, and taking the data points corresponding to the difference between the sliding window size and the prediction time scale as the time data input sequence.

[0069] S103. Reduce the dimension of the image sequence based on a 3D convolutional neural network, and splice the dimension-reduced image sequence with the photovoltaic power generation data to form a time feature sequence;

[0070] See Figure 3, dimensionality reduction is performed on the image sequence based on a 3D convolutional neural network, specifically: reducing the image sequence data with the specification of L*C*H*W to the specification of L*N, where L is the length of the image sequence, C is the number of image channels, H is the image height, W is the image width, and N is the number of extracted features.

[0071] See Figure 4 , the 3D convolutional neural network includes four consecutive 3D convolutional layers, and a ReLU activation function and a Dropout layer are connected after each 3D convolutional layer; the 3D convolutional layers include: Conv1, Conv2, Conv3, and Conv4; the 3D convolutional layers reduce the image sequence to a feature sequence with 8 parameters;

[0072] Conv1 has an input channel number of 3, an output channel number of 32, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0073] Conv2 has an input channel number of 32, an output channel number of 16, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0074] Conv3 has an input channel number of 16, an output channel number of 8, a convolutional kernel size of (3, 4, 4), a stride of (1, 4, 4), and a padding of (1, 0, 0);

[0075] Conv4 has an input channel number of 8, an output channel number of 8, a convolutional kernel size of (3, 2, 2), a stride of (1, 2, 2), and a padding of (1, 0, 0).

[0076] S104, input the time feature sequence into the DLinear structure for decomposition to obtain a trend sequence and a residual sequence; and in the DLinear structure, construct a fully connected relationship between the trend sequence, the residual sequence, and the power generation power sequence to be predicted, so that the power generation power sequence to be predicted is directly output through the decomposed multiple sequences.

[0077] Input the time feature sequence into the DLinear structure, and each variable time series is decomposed into a trend sequence and a residual sequence. The sequence part calculated by moving average represents the relatively smooth part in the data and is used as the long-term trend of the time series; the residual sequence is the rapidly changing component and noise.

[0078] See Figure 5, the time feature sequence is a time series with 9 original feature parameters; when the time feature sequence is decomposed into a trend sequence and a residual sequence, each original time series will generate two new sequences, a trend sequence and a residual sequence; for 9 original time series, 2 * 9 = 18 new sequences will be obtained; a fully connected relationship is established between the 18 newly generated sequences and the power generation power sequence to be finally predicted; the established fully connected relationship connects the 18 * L parameters of the 18 new sequences with the respective parameters on the power generation power sequence to be predicted according to the randomly assigned initial weights; that is, the parameter points at each moment of each new sequence are connected to the parameter points at each moment of the power generation power sequence to be predicted; the 18 * L parameter points of the new sequence are mapped to the L parameter points of the power generation power sequence to be predicted through a fully connected layer; the parameter points at each moment of the power generation power sequence to be predicted are the weighted sum of the parameter points at all moments of the new sequence; calculate the mean square error MSE between the power generation power sequence to be predicted and the historical power generation power sequence; calculate the gradient of the loss function with respect to the weights through the backpropagation algorithm; use an optimizer to update the weights according to the gradient; repeatedly perform forward propagation, loss calculation, backpropagation, and weight update until the model converges or reaches a predetermined number of training epochs; and adaptively adjust the weight distribution during iterative training, and the trained model directly outputs the power generation power sequence of photovoltaic power generation to be predicted.

[0079] See Figure 2 , the present invention discloses a short-term photovoltaic power generation power time series prediction system for household energy, including:

[0080] An acquisition module, which acquires ground-based cloud map data and photovoltaic power generation data;

[0081] A preprocessing module, which preprocesses the acquired photovoltaic power generation data, and based on the time data input sequence of the preprocessed photovoltaic power generation data, obtains an image sequence matching the photovoltaic power generation data in the ground-based cloud map data;

[0082] A dimensionality reduction module, which reduces the dimension of the image sequence based on a 3D convolutional neural network, and splices the dimension-reduced image sequence with the photovoltaic power generation data to form a time feature sequence;

[0083] A decomposition module, which inputs the time feature sequence into a DLinear structure for decomposition to obtain a trend sequence and a residual sequence, and in the DLinear structure, constructs a fully connected relationship between the trend sequence, the residual sequence and the power generation power sequence to be predicted, so that it directly outputs the power generation power sequence to be predicted through the multiple sequences after decomposition.

[0084] Embodiment:

[0085] The present invention discloses a short-term photovoltaic power time series prediction method for household energy, including:

[0086] Collect data for the household energy system to be predicted, including installing a photovoltaic power self-recording meter and a fish-eye camera for obtaining ground-based cloud map data. The photovoltaic power data is measured with a 5-minute time accuracy. The fish-eye camera selects an ordinary household fish-eye camera. After certain light-shielding and waterproof protection operations, it is installed vertically and reversely beside the photovoltaic panel at a position about 1 m above the ground. After arranging the instruments, data collection is carried out for a period of time. The fish-eye camera uses a 1 / 2.5-inch CMOS, with a minimum illuminance of 0.1 Lux in color mode, fixed-focus shooting, fixed aperture, a working environment of -30°C - 60°C, and a size of 134 mm × 134 mm × 51 mm.

[0087] After collecting data for a period of time, first perform preprocessing operations on the data. For the photovoltaic power data, first perform outlier processing, detect and remove outliers by using the 3σ rule to ensure the reliability of the data. Secondly, fill in the missing values, and use the interpolation method to fill in the missing parts of the data to maintain the continuity of the data.

[0088] For the ground-based cloud map data, determine the time range of the image sequence to be matched according to the time stamps of the time data input sequence; extract key frames from the images in the ground-based cloud map dataset at fixed frame intervals and save them in image format; sort the extracted images according to the time stamps and screen out the image sequence corresponding to the time range of the time data input sequence; transform the screened image sequence based on the nearest neighbor interpolation method to uniformly transform it into a data size of 3*128*128, which is convenient for model loading and use, and at the same time reduces the storage space occupation.

[0089] After completing the data preprocessing, form a training set of time series model input data. The present invention is applicable to time series prediction tasks with four prediction time scales (0.5 hour, 1 hour, 1.5 hours, and 2 hours). For the dataset, intercept 10 hours of data (120 data points per day) from 7:00 to 17:00 every day for model training, and adopt the sliding window method with a fixed window size of 4 hours (48 data points). The model input time lengths corresponding to different prediction time scales are: 3.5 hours (predicting 0.5 hour, 42 data points predicting 6 data points), 3 hours (predicting 1 hour, 36 data points predicting 12 data points), 2.5 hours (predicting 1.5 hours, 30 data points predicting 18 data points), and 2 hours (predicting 2 hours, 24 data points predicting 24 data points). For each input sequence, simultaneously match the corresponding image sequence in the ground-based cloud map dataset to jointly constitute the input part of the model.

[0090] Reduce the image sequence data with the specification of L*C*H*W to the specification of L*N, where L is the length of the image sequence, C is the number of image channels, H is the image height, W is the image width, and N is the number of extracted features.

[0091] The 3D convolutional neural network contains four consecutive 3D convolutional layers, and a ReLU activation function and a Dropout layer are connected after each 3D convolutional layer; in order to enhance the non-linear expression ability of the network, a ReLU activation function is connected after each convolutional layer, and at the same time, a Dropout layer (with a probability of 0.3) is introduced to reduce the risk of overfitting. After a series of operations, the input image sequence data with the specification of L*C*H*W is dimensionally reduced to L*8*1*1. Finally, the model extracts 8 feature sequences from the image through dimensionality reduction as the input of the time series model.

[0092] After completing the above tasks, concatenate the 8 feature sequences after dimensionality reduction of the image sequence with the historical photovoltaic power generation at the corresponding time to form a time series of 9 feature parameters. Input the latest feature sequence into the DLinear structure, and each variable time series is decomposed into a trend series and a remainder series. The sequence part calculated by moving average represents the smoother part of the data and serves as the long-term trend of the time series. Secondly, the result of subtracting the trend series from the original sequence represents the remaining part, that is, the fast-changing components and noise. The DLinear structure part establishes a fully connected relationship between 2*9 sequences and the power generation prediction sequence to be predicted, so that each component directly contributes to the final power generation prediction. DLinear includes decomposing to obtain the trend / remainder sequence and then predicting the power generation sequence in a fully connected form; for example, for 4h of photovoltaic data, predict the photovoltaic power generation data for the next 1h through the image sequence and photovoltaic historical data for the previous 3h. Then, a fully connected relationship is required between the trend sequence, remainder sequence decomposed from the image sequence and photovoltaic historical data for the previous 3h and the power generation sequence for the next 1h.

[0093] During the backpropagation process of the model, the parameters of the 3D-CNN module and the DLinear module share the same computational graph. While the time series model updates its parameters, the structural parameters of the 3D-CNN module are also synchronously optimized based on the unified gradient. This collaborative optimization mechanism significantly improves the model's ability to model the time series relationship of photovoltaic power generation.

[0094] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting short-term photovoltaic power generation time series of household energy, characterized in that: include: Collect ground-based cloud map data and photovoltaic power generation data; Preprocessing the collected photovoltaic power generation data, and obtaining an image sequence matching the photovoltaic power generation data in the ground-based cloud image data based on the time data input sequence of the preprocessed photovoltaic power generation data; The image sequence is reduced in dimension based on a 3D convolutional neural network, and the reduced-dimensional image sequence is spliced ​​with photovoltaic power generation data to form a time feature sequence; The time feature sequence is input into the DLinear structure for decomposition to obtain the trend sequence and the residual sequence. In the DLinear structure, a fully connected relationship between the trend sequence, the residual sequence and the power generation sequence to be predicted is constructed, so that the power generation sequence to be predicted can be directly output through the decomposed multiple sequences.

2. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 1 is characterized in that: The collected photovoltaic power generation data is preprocessed as follows: The outliers of photovoltaic power generation data are detected and eliminated based on the 3σ rule, and missing values ​​are filled by linear interpolation. The ground-based cloud map data is obtained by photographing the sky cloud map with a household fisheye camera.

3. The method for predicting the short-term photovoltaic power generation power of household energy according to claim 2 is characterized in that: The time data input sequence of the pre-processed photovoltaic power generation data is used to obtain an image sequence matching the photovoltaic power generation data in the ground-based cloud image data, specifically: Determine the time range of the image sequence to be matched according to the timestamp of the time data input sequence; Extract key frames from the images in the ground-based cloud image dataset at fixed frame intervals and save them in image format; Sort the extracted images by timestamps and filter out the image sequences corresponding to the time range of the time data input sequence; The filtered image sequence is transformed based on the nearest neighbor interpolation method to obtain images of uniform size.

4. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 3 is characterized in that: The time data input sequence is specifically: Intercept a fixed time period, obtain the amount of data contained in the fixed time period, and filter out the amount of data corresponding to the window size from the fixed time period through a sliding window; Based on the set prediction time scale, construct the corresponding time data input sequence; Among them, the corresponding data input sequence is constructed as follows: the data points corresponding to the prediction time scale are taken as the quantity to be predicted, and the data points corresponding to the difference between the sliding window size and the prediction time scale are taken as the time data input sequence.

5. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 4 is characterized in that: The dimensionality reduction of the image sequence based on the 3D convolutional neural network is specifically as follows: reducing the image sequence data with a specification of L*C*H*W to a specification of L*N, wherein L is the image sequence length, C is the number of image channels, H is the image height, W is the image width, and N is the number of extracted features.

6. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 5 is characterized in that: The 3D convolutional neural network includes four consecutive 3D convolutional layers, each of which is connected to a ReLU activation function and a Dropout layer; the 3D convolutional layers include: Conv1, Conv2, Conv3 and Conv4; the 3D convolutional layers reduce the dimension of the image sequence to a feature sequence of 8 parameters; Conv1 has 3 input channels, 32 output channels, a convolution kernel size of (3,4,4), a stride of (1,4,4), and padding of (1,0,0); Conv2 has 32 input channels, 16 output channels, a convolution kernel size of (3,4,4), a stride of (1,4,4), and padding of (1,0,0); Conv3 has 16 input channels, 8 output channels, convolution kernel size (3,4,4), stride (1,4,4), and padding (1,0,0). The Conv4 has 8 input channels, 8 output channels, a convolution kernel size of (3,2,2), a stride of (1,2,2), and padding of (1,0,0).

7. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 6, characterized in that: The time feature sequence is input into the DLinear structure for decomposition to obtain a trend sequence and a residual sequence. Specifically, the time feature sequence is input into the DLinear structure, and each variable time series is decomposed into a trend sequence and a residual sequence. The sequence part calculated by the sliding average represents the smoother part of the data, which serves as the long-term trend of the time series; the residual sequence is a rapidly changing component and noise.

8. The method for predicting the short-term photovoltaic power generation time series of household energy according to claim 7, characterized in that: In the DLinear structure, a fully connected relationship is constructed between the trend sequence, the residual sequence and the power generation sequence to be predicted, so that the power generation sequence to be predicted is directly output through the decomposed multiple sequences, specifically: the time feature sequence has a time series of 9 original feature parameters; when the time feature sequence is decomposed into a trend sequence and a residual sequence, each original time series will generate two new sequences, a trend sequence and a residual sequence; for 9 original time series, 2*9=18 new sequences will be obtained; a fully connected relationship is established between the 18 newly generated sequences and the final power generation sequence to be predicted; the established fully connected relationship connects the 18*L parameters of the 18 new sequences with the various parameters on the power generation sequence to be predicted according to the randomly assigned initial weights; that is, the parameter points of each new sequence at each moment are connected with the parameter points of the power generation sequence to be predicted at each moment; The 18*L parameter points of the new sequence are mapped to the L parameter points of the power generation sequence to be predicted through a fully connected layer; the parameter point of the power generation sequence to be predicted at each moment is the weighted sum of the parameter points of the new sequence at all moments; the mean square error (MSE) between the power generation sequence to be predicted and the historical power generation sequence is calculated; the gradient of the loss function to the weight is calculated through the back propagation algorithm; the weight is updated according to the gradient using the optimizer; forward propagation, loss calculation, back propagation and weight update are repeated until the model converges or the predetermined number of training rounds is reached; The weight distribution is adaptively adjusted during iterative training, and the trained model directly outputs the predicted photovoltaic power generation sequence.

9. A short-term photovoltaic power generation time series prediction system for household energy, characterized in that: include: A collection module, wherein the collection module collects ground cloud map data and photovoltaic power generation data; A preprocessing module, which preprocesses the collected photovoltaic power generation data, and obtains an image sequence matching the photovoltaic power generation data in the ground-based cloud image data based on a time data input sequence of the preprocessed photovoltaic power generation data; A dimensionality reduction module, which reduces the dimensionality of the image sequence based on a 3D convolutional neural network, and splices the image sequence after dimensionality reduction with the photovoltaic power generation data to form a time feature sequence; A decomposition module, wherein the decomposition module inputs the time feature sequence into the DLinear structure for decomposition, obtains a trend sequence and a residual sequence, and constructs a fully connected relationship among the trend sequence, the residual sequence and the power generation sequence to be predicted in the DLinear structure, so that the power generation sequence to be predicted can be directly output through the decomposed multiple sequences.

Images