Day-ahead photovoltaic power generation prediction method and system based on seasonal decomposition and convolutional network

Abstract

The application relates to a next-day photovoltaic power generation prediction method and system based on seasonal decomposition and a convolution network, and the method comprises the following steps: S1, collecting historical measured meteorological data and power generation power in a photovoltaic power generation process, combining meteorological prediction provided by a meteorological station, and respectively constructing historical measured data and meteorological prediction data; S2, performing seasonal sequence decomposition on the historical measured data and the meteorological prediction data to obtain a historical measured seasonal trend sequence, a historical measured fluctuation sequence and a meteorological prediction fluctuation sequence; S3, inputting the historical measured fluctuation sequence and the meteorological prediction fluctuation sequence into a multi-core gated convolution neural network to obtain high-level meteorological representation; S4, inputting the high-level meteorological representation into a time-specific prediction network to calculate a predicted fluctuation sequence of photovoltaic power generation power in a future day; and S5, adding the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain a photovoltaic power prediction result. The method provided by the application improves the accuracy of photovoltaic power prediction.

Description

Technical Field

[0001] This invention relates to the fields of new energy power generation, time series forecasting, and data mining, specifically to a method and system for forecasting next-day photovoltaic power generation based on seasonal decomposition and convolutional networks. Background Technology

[0002] In today's era, solar power generation is attracting increasing attention due to its pollution-free nature and unlimited resources. However, photovoltaic (PV) power generation is highly unstable, as it is heavily dependent on weather conditions. This volatility poses challenging problems for grid system scheduling, power balancing, and power distribution. Therefore, accurate PV power generation forecasting is crucial to mitigating the detrimental effects of these issues.

[0003] During the operation of a photovoltaic (PV) system, various operational data are typically collected, including power generation records and weather observations. When conditions permit, local weather forecasts are also acquired to assist in prediction. By analyzing this data, future power generation can be predicted. Depending on actual needs, PV power generation forecasts can range from a few minutes to several months. Forecasts lasting a few minutes primarily serve the control of individual generating units; forecasts lasting several hours (or days) primarily serve the day-ahead dispatching of the power grid system and electricity market transactions; and forecasts lasting several months or longer are mainly used for power system construction planning.

[0004] Because weather forecast data may be missing, engineers need to predict future power generation based on historical power generation data and weather observation records. Many studies use recorded weather observations as input for power forecasting tasks. One assumption in these studies is that future irradiance and photovoltaic power generation are related to current weather. They employ various methods, such as neural networks, heterogeneous regression, and deep belief networks. When the time span of recorded weather observations expands, time series analysis methods are used, such as autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), and several variants of recurrent neural networks (RNN), long short-term memory networks (LSTM), and convolutional neural networks (CNN). Recently, methods based on self-attention mechanisms have emerged, significantly improving the predictive power of long-term series. These studies demonstrate good predictive performance.

[0005] Furthermore, due to the strong correlation between weather forecasts and photovoltaic (PV) power generation, an increasing number of studies are using weather forecasts as a primary source of input data. Based on the predicted weather, future PV power generation is predicted using physical models, machine learning, and artificial neural networks. However, weather-based forecasting methods still have some limitations. First, since the purpose of weather forecasts is not solely for PV power generation prediction, the models trained on weather forecasts may not be optimally suited for predicting features related to PV power generation. Second, weather forecasts are typically conservative, and large weather changes may not be accurately predicted. Finally, due to limited computing power, weather forecasts usually only cover a few locations, which may differ from the location of the PV system, leading to forecast errors. Therefore, improving the robustness of weather forecasts in the PV power generation field, thereby increasing the accuracy of PV power prediction, has become a pressing issue. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for predicting next-day photovoltaic power generation based on seasonal decomposition and convolutional networks.

[0007] The technical solution of this invention is: a method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks, comprising:

[0008] Step S1: Collect historical measured meteorological data and power generation during the photovoltaic power generation process, combine them with meteorological forecasts provided by the meteorological station, and perform preprocessing to construct historical measured data and meteorological forecast data respectively;

[0009] Step S2: Perform seasonal sequence decomposition on the historical measured data and the meteorological forecast data respectively to obtain the historical measured seasonal trend sequence and the historical measured fluctuation sequence, as well as the meteorological forecast fluctuation sequence.

[0010] Step S3: Input the historical measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain advanced meteorological characterization;

[0011] Step S4: Input the advanced meteorological characterization into a time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day;

[0012] Step S5: Add the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain the final photovoltaic power prediction result.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] 1. This invention discloses a method for predicting next-day photovoltaic (PV) power generation based on seasonal decomposition and convolutional networks. The method employs seasonal sequence decomposition to predict next-day PV power output. Through seasonal sequence decomposition, a seasonal trend sequence and a deviation sequence are obtained. The seasonal trend sequence retains the seasonal variation trend of PV power generation, while the deviation sequence contains more information specific to the current day. This seasonal sequence decomposition can eliminate seasonal differences, thus ensuring consistent data distribution across different seasons.

[0015] 2. This invention constructs a multi-core gated convolutional neural network. The network's convolution-based encoder encodes the measured weather of the previous day and the predicted weather of the next day to learn short-term temporal correlations. The convolution-based decoder combines the two encodings of the last day's observation and the next day's prediction to learn cross-day temporal correlations.

[0016] 3. This invention utilizes a multi-kernel gated convolutional neural network to achieve more accurate and robust learning of local cross features between time steps. It uses multiple convolutional kernels with different receptive fields to learn different local cross features between time steps, and employs a single kernel to further extract multiple temporal correlations.

[0017] 4. This invention utilizes a time-specific prediction network to make more accurate predictions. Because the correlation between different features and photovoltaic power generation varies greatly at different times, different prediction parameters are used for photovoltaic power prediction at different times. Attached Figure Description

[0018] Figure 1 This is a flowchart of a next-day photovoltaic power generation prediction method based on seasonal decomposition and convolutional networks in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the MGCNN structure in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the encoder 1 in an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of the encoder 2 in an embodiment of the present invention;

[0022] Figure 5 This is a schematic diagram of the decoder structure in an embodiment of the present invention;

[0023] Figure 6 This is a schematic diagram of the prediction of a time-specific prediction network in an embodiment of the present invention;

[0024] Figure 7 This is a schematic diagram of the photovoltaic power generation prediction method for the next day in an embodiment of the present invention;

[0025] Figure 8 This is a structural block diagram of a next-day photovoltaic power generation prediction system based on seasonal decomposition and convolutional networks, as described in an embodiment of the present invention. Detailed Implementation

[0026] This invention provides a method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks. By decomposing the seasonal sequence, the method avoids the impact of seasonal fluctuations on data distribution. Furthermore, it learns the time dependence between historical measured data and meteorological forecast data through a multi-core gated convolutional neural network. Finally, it employs a time-specific power prediction network to improve the accuracy of photovoltaic power prediction.

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below through specific implementations and in conjunction with the accompanying drawings.

[0028] Example 1

[0029] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks, comprising the following steps:

[0030] Step S1: Collect historical measured meteorological data and power generation during the photovoltaic power generation process, combine them with meteorological forecasts provided by the meteorological station, and perform preprocessing to construct historical measured data and meteorological forecast data respectively;

[0031] Step S2: Perform seasonal series decomposition on the historical measured data and the meteorological forecast data respectively to obtain the historical measured seasonal trend series and the historical measured fluctuation series, as well as the meteorological forecast fluctuation series.

[0032] Step S3: Input the historical measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain advanced meteorological characterization;

[0033] Step S4: Input the advanced meteorological characterization into the time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day;

[0034] Step S5: Add the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain the final photovoltaic power prediction result.

[0035] In one embodiment, step S1 above involves collecting historical measured weather data and power generation during the photovoltaic power generation process, combining this with weather forecasts provided by the meteorological station, and performing preprocessing to construct historical measured data and weather forecast data, specifically including:

[0036] Step S11: Collect measured meteorological data and power generation during the photovoltaic power generation process within the i-th day. The measured meteorological data includes five characteristic values ​​measured every 15 minutes: irradiance, temperature, humidity, air pressure and wind speed. The power generation includes one characteristic value every 15 minutes: power generation of the power station.

[0037] Constructing historical measured data x from day i before i ={x i,1 ,x i,2 ,x i,3 ,…x i,96}; where x i,t ∈R 6 Let t represent the historical measured data at time t on the i-th day, where 1 ≤ t ≤ 96; since data is collected every 15 minutes, the length of the data collected in 24 hours is 96.

[0038] Step S12: The historical measured data is processed using a normalization method to scale the numerical values, adjusting the range to [0,1], and constructing the historical measured data X = {x1,x2,x3,…,x...} for the previous 15 days. 15};

[0039] Step S13: Collect photovoltaic weather forecast data for the next day. The forecast includes five characteristic values ​​predicted every 15 minutes: irradiance, temperature, humidity, air pressure, and wind speed.

[0040] Constructing next day's weather forecast data y i ={y i,1 ,y i,2 ,y i,3 ,…,y i,96}, where y i,t ∈R 5 This indicates that for the predicted data at time t on the i-th day prior, 1 ≤ t ≤ 96; since a weather forecast is provided every 15 minutes, the length of the weather forecast for a 24-hour day is 96.

[0041] Step S13: Use a normalization method to scale the numerical scale of the next day's weather forecast data, adjusting the numerical range to [0,1]; construct the forecast data for the previous 14 days and the weather forecast data for the day to be predicted, Y = {y0,y1,y2,…,y...} 14}, where y0 is the predicted data for the date to be predicted.

[0042] In one embodiment, step S2 above, which involves performing seasonal series decomposition on the historical measured data and the meteorological forecast data to obtain the historical measured seasonal trend series and the historical measured fluctuation series, as well as the meteorological forecast fluctuation series, specifically includes:

[0043] Step S21: Calculate the mean of the historical measured data X at each time point 1 ≤ t ≤ 96: Obtain the historically measured seasonal trend sequence: The latest historical measured data is the observation data from one day prior, x1={x 1,1 ,x 1,2 ,x 1,3 ,…x 1,96 The historically measured fluctuation sequence is obtained by subtracting the historically measured seasonal trend sequence from this sequence.

[0044]

[0045] Step S22: Calculate the mean of the meteorological forecast data Y at each time point 1 ≤ t ≤ 96: Obtain the seasonal trend series of weather forecasts The latest meteorological forecast data is the observation data y0={y 0,1 ,y 0,2 ,y 0,3 ,…,y 0,96 The fluctuation sequence of the weather forecast is obtained by subtracting the seasonal trend sequence of the weather forecast from this sequence:

[0046]

[0047] In one embodiment, step S3 above, which involves inputting the historically measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain a high-level meteorological representation, specifically includes:

[0048] Step S31: Construct a multi-kernel gated convolutional neural network, including: encoder 1, encoder 2 and a decoder; wherein each encoder and decoder contains two layers of multi-kernel gated convolutional MGCNN, and each MGCNN contains two convolutional layers and corresponding gated convolutional layers;

[0049] First, the first layer of MGCNN is a multi-kernel convolutional layer, for the input... We use N convolutional kernels, each with an input dimension of d0, an output dimension of d1, and a kernel size of 3; the operation of convolutional kernel i is shown in the following formula:

[0050]

[0051] a i,t =g(W i*i X+b i )

[0052] Where conv(·) represents a single convolution kernel, and the output of convolution kernel i is: These are the corresponding convolution kernel parameters, *i This represents a convolution operation with an expansion coefficient of i, where i = 1, 2, ..., N, and g is the Leaky ReLU activation function, as shown in the following equation:

[0053] g(x) = max(0,x) + l·min(x,0)

[0054] Where l is a very small positive number;

[0055] The output result of the convolution kernel is a. i By concatenating the results, we obtain the output A of the first convolutional layer:

[0056] A = concat(a1, a2, a3, ... a i …a N )

[0057] Here, `concat` is the concatenation operation.

[0058] Next, A is input into the second single-kernel convolutional layer, with an output dimension of d2 and an inflation factor of 1. The output of this convolution is as follows:

[0059] C = conv(A)

[0060] in,

[0061] For simplicity, all the above formulas are simplified into a single operator, multi(·), i.e.:

[0062] C=multi(X)=conv(concat(conv1(X),conv2(X),…,conv N (X))

[0063] Following the structure of C, construct two parallel convolution channels:

[0064] C1 = multi1(X)

[0065] C2 = multi2(X)

[0066] The output of the gated convolutional layer is as follows:

[0067] H = MGCNN(X) = C1⊙σ(C2)

[0068] Where σ(·) is the Sigmoid activation function, ⊙ is the element-wise multiplication operation, and MGCNN is the multi-kernel gated convolution;

[0069] For simplicity, the above complete multi-core gated convolution is represented as:

[0070] H = MGCNN(X);

[0071] like Figure 2 The diagram shows the structure of MGCNN, including two parallel convolutional channels C1 and C2. Each convolutional channel contains two convolutional layers: the first layer is a multi-kernel convolutional layer, and the second layer is a single-kernel convolutional layer.

[0072] In this embodiment of the invention, through verification, it is found that using 4 convolutional kernels in the first multi-kernel convolutional layer of MGCNN, with each convolutional kernel having an output dimension of 6, and using a single-kernel convolutional layer in the second layer with an output dimension of 24, achieves better results, i.e., N=4, d1=6, d2=24.

[0073] Step S32: Place X d Input encoder 1, after passing through the first layer of MGCNN, outputs H x,1 :

[0074] H x,1 =MGCNN(X) d )

[0075] Among them, X d The input dimension is d0;

[0076] After passing through the second MGCNN layer, encoder 1 outputs H. x,2 :

[0077] H x,2 =MGCNN(H x,1 )+H x,1

[0078] Let H x =H x,2 As a volatility encoding of historical measured data;

[0079] like Figure 3 The diagram shown is a schematic of encoder 1.

[0080] Step S33: Y d Input encoder 2, after passing through the first layer of MGCNN, outputs H y,1 :

[0081] H y,1 =MGCNN(Y d )

[0082] After passing through the second MGCNN layer, the output of encoder 2 is:

[0083] H y,2 =MGCNN(H y,1 )+H y,1

[0084] Let H y =Hy,2 As a way to encode the volatility of meteorological forecast data;

[0085] like Figure 4 The diagram shown is a structural schematic of encoder 2;

[0086] Step S34: H x and H y The input decoder first fuses the two data points. For times 1 ≤ t ≤ 96, the fused wave behavior code H is calculated.

[0087]

[0088] in, concat represents the concatenation operation;

[0089] The volatility code H based on the historical measured data obtained from the aforementioned steps x Volatility coding of meteorological forecast data H y In this step, the two are concatenated moment by moment. This is because the volatility coding of historical measured data shows the fluctuation of each moment on the day preceding the forecast date relative to the measured value at the same time of the current season; while the volatility coding of meteorological forecast data shows the fluctuation of each moment on the forecast date relative to the predicted value at the same time of the current season. Therefore, the volatility coding of historical observation data at the same time is concatenated with the volatility coding of meteorological forecast data to obtain the fused volatility coding for the corresponding moment.

[0090] Step S35: Pass the fused volatility code H through two layers of MGCNN to obtain a high-level meteorological characterization. The formula for calculating M is as follows:

[0091] M = MGCNN(MGCNN(H) + H) y ).

[0092] like Figure 5 The diagram shown is a schematic of the decoder structure.

[0093] In one embodiment, step S4 above, which involves inputting advanced meteorological characterization into a time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day, specifically includes:

[0094] By inputting the advanced meteorological characterization M into a time-specific prediction network, the predicted fluctuation sequence of photovoltaic power generation within the next day is calculated according to the following formula:

[0095]

[0096] Among them, M t The higher-level meteorological characterization at time t is M = {M1, M2, M3, ..., M}. 96The t-th line in}; and b t ∈R is a specific parameter; This represents the predicted fluctuation value of photovoltaic power generation at time t;

[0097] For all times of the day, 1 ≤ t ≤ 96, using specific prediction parameters, the predicted power fluctuation value for that day is... for:

[0098]

[0099] like Figure 6 The diagram shown is a prediction schematic of a time-specific prediction network.

[0100] In one embodiment, step S5 above—adding the predicted fluctuation sequence to the historically measured seasonal trend sequence to obtain the final photovoltaic power prediction result—specifically includes:

[0101] Step S51: From historically measured seasonal trend sequences: Obtain the seasonal trend sequence of power generation corresponding to the power generation characteristics of the power plants.

[0102] Step S52: Put p m and By summing the results at each time step, the final power prediction for the next day can be obtained. for:

[0103]

[0104] Step S53: Construct the L2 loss function as the loss function for photovoltaic power prediction. The calculation formula is as follows:

[0105]

[0106] in, Let p be the predicted photovoltaic power at time t. t Let t be the actual value of photovoltaic power at time t.

[0107] In this embodiment of the invention, a multi-core gated convolutional neural network is trained using the backpropagation algorithm, and the adaptive motion estimation algorithm (Adam) is selected as the optimization algorithm. The initial learning rate is 0.0005 and the batch size is 32.

[0108] Figure 7 A schematic diagram illustrating the method for predicting photovoltaic power generation the following day is shown.

[0109] After the prediction model completes its predictions, the root square error (RMSE) and mean absolute error (MAE) are selected as statistical performance indicators for the prediction method. Here, the smaller the values ​​of RMSE and MAE, the better the model's estimation performance. The expressions for MAE and RMSE are shown in the following formulas:

[0110]

[0111]

[0112] Where T is the length of the data. To make the results more intuitive, a normalization method was used for the power to be predicted, limiting the value to the range of [0,1]. From the normalized error results, the proportion of the deviation value relative to the installed capacity of the photovoltaic power station can be observed. Power generation data from three photovoltaic power stations in Anhui Province (Dongdawei Photovoltaic Power Station, Zaoshan Photovoltaic Power Station, and Bohu Photovoltaic Power Station) were collected for algorithm verification. Decision trees, linear regression, K-nearest neighbors, random forest, AdaBoost, XGBoost, Transformer, and Autoformer were selected as control methods. The prediction results are compared in Table 1. In the prediction of the three stations, the method of this invention achieved the lowest MAE and RMSE, i.e., the best performance.

[0113] Table 1 Comparison of Prediction Results

[0114]

[0115] This invention discloses a method for predicting next-day photovoltaic (PV) power generation based on seasonal decomposition and convolutional networks. The method employs seasonal sequence decomposition to predict next-day PV power output. Through seasonal sequence decomposition, a seasonal trend sequence and a deviation sequence are obtained. The seasonal trend sequence retains the seasonal variation trend of PV power generation, while the deviation sequence contains more day-specific information. This seasonal sequence decomposition eliminates seasonal differences, thus maintaining a consistent data distribution across different seasons. This invention constructs a multi-kernel gated convolutional neural network. The network's convolution-based encoder encodes the previous day's measured weather and the next day's predicted weather to learn short-term temporal correlations; while the decoder combines the two encodings—the last day's observations and the next day's predictions—to learn cross-day temporal correlations. This invention utilizes a multi-kernel gated convolutional neural network to achieve more accurate and robust learning of local cross-features between time steps. Multiple convolutional kernels with different receptive fields are used to learn different local cross-features between time steps, and a single kernel is used to further extract multiple temporal correlations. This invention utilizes a time-specific prediction network to make more accurate predictions. Because the correlation between different characteristics and photovoltaic power generation varies greatly at different times, different prediction parameters are used to predict photovoltaic power at different times.

[0116] Example 2

[0117] like Figure 8 As shown, this embodiment of the invention provides a next-day photovoltaic power generation prediction system based on seasonal decomposition and convolutional networks, including the following modules:

[0118] A historical measured data and meteorological forecast data module 61 is constructed to collect historical measured meteorological data and power generation during the photovoltaic power generation process. Combined with the meteorological forecast provided by the meteorological station, and after preprocessing, historical measured data and meteorological forecast data are constructed respectively.

[0119] A module 62 for constructing historical measured fluctuation sequences and meteorological forecast fluctuation sequences is used to perform seasonal sequence decomposition on historical measured data and meteorological forecast data respectively, so as to obtain historical measured seasonal trend sequences and historical measured fluctuation sequences, as well as meteorological forecast fluctuation sequences.

[0120] The advanced meteorological characterization module 63 is used to: input historical measured fluctuation sequences and meteorological forecast fluctuation sequences into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain advanced meteorological characterizations;

[0121] The photovoltaic power generation prediction fluctuation sequence module 64 is used to input advanced meteorological characteristics into a time-specific prediction network to calculate the prediction fluctuation sequence of photovoltaic power generation within the next day.

[0122] The photovoltaic power prediction module 65 is used to add the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain the final photovoltaic power prediction result.

[0123] The above embodiments are provided merely for the purpose of describing the present invention and are not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the invention should be covered within the scope of the invention.

Claims

1. A method for predicting next-day photovoltaic power generation based on seasonal decomposition and convolutional networks, characterized in that, include: Step S1: Collect historical measured weather data and power generation during the photovoltaic power generation process, combine them with weather forecasts provided by the meteorological station, and perform preprocessing to construct historical measured data and weather forecast data respectively, specifically including: Step S11: Collect the previous... The measured meteorological data and power generation during the photovoltaic power generation process within the day, wherein the measured meteorological data includes five characteristic values ​​measured every 15 minutes: irradiance, temperature, humidity, air pressure and wind speed, and the power generation includes one characteristic value every 15 minutes: power generation of the power station. Constructing the first Historical measured data of Japan ;in Indicates the previous number day Historical measured data at any given time ; Step S12: The historical measured data is processed using a normalization method to scale the numerical values, adjusting the range to [0,1], and constructing the historical measured data for the previous 15 days. ; Step S13: Collect photovoltaic weather forecast data for the next day, which includes five characteristic values ​​predicted every 15 minutes: irradiance, temperature, humidity, air pressure, and wind speed. Constructing next day's weather forecast data ,in Indicates the number of steps forward. day Predicted data for each moment, ; Step S14: The numerical scale of the next day's weather forecast data is scaled down using a normalization method, adjusting the numerical range to [0,1]; the forecast data for the previous 14 days and the weather forecast data for the day to be predicted are constructed. ,in This is the forecast data for the date to be predicted; Step S2: Perform seasonal series decomposition on the historical measured data and the meteorological forecast data respectively to obtain the historical measured seasonal trend series and the historical measured fluctuation series, as well as the meteorological forecast fluctuation series, specifically including: Step S21: Calculate the historical measured data In each Mean at time: This yields a historically measured seasonal trend sequence: The latest historical measurement data is the observation data from one day prior. The historically measured fluctuation sequence is obtained by subtracting the historically measured seasonal trend sequence from this sequence: Step S22: Calculate the meteorological forecast data. In each Mean at time: The seasonal trend series of meteorological forecasts were obtained. The latest weather forecast data is based on observational data for the day to be forecasted. The fluctuation sequence of the weather forecast is obtained by subtracting the seasonal trend sequence of the weather forecast from this sequence: ； Step S3: Input the historical measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain advanced meteorological characterization; Step S4: Input the advanced meteorological characterization into a time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day; Step S5: Add the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain the final photovoltaic power prediction result.

2. The method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks according to claim 1, characterized in that, Step S3: Input the historically measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain a high-level meteorological representation, specifically including: Step S31: Construct a multi-kernel gated convolutional neural network, including: encoder 1, encoder 2 and a decoder; wherein each encoder and decoder contains two layers of multi-kernel gated convolutions. each It contains two convolutional layers and corresponding gated convolutional layers; first, The first layer is a multi-kernel convolutional layer, for the input ,use There are n convolutional kernels, each with an input dimension of 1. The output dimension is The kernel size is 3; the convolution kernel The operation is shown in the following formula: in, Represents a single convolution kernel. The output result is ; These are the corresponding convolution kernel parameters. Indicates the coefficient of thermal expansion is Convolution operation, , It is the LeakyReLU activation function, as shown in the following equation: in, It is a very small positive number; Output the convolution kernel. The outputs of the first convolutional layer are concatenated to obtain the output of the convolutional layer. : in, It's a splicing operation. ; Secondly, The input is the second single-kernel convolutional layer, and the output dimension is... With an inflation factor of 1, the output of this convolution is as follows: in, ; For simplicity, all the above formulas are simplified into a single operator. ,Right now: according to The structure is used to construct two parallel convolution channels: The output of the gated convolutional layer is as follows: in, It is the Sigmoid activation function. This is an element-wise multiplication operation. Multi-core gated convolution; For simplicity, the above complete multi-core gated convolution is represented as: ； Step S32: The input is to encoder 1, passing through the first layer. The output is : in, Input dimension is ; After passing through the second layer Then the encoder 1 outputs: : make As a volatility encoding of the historical measured data; Step S33: ... The input is to the encoder 2, passing through the first layer. The output is : After passing through the second layer Then the output of encoder 2 is: make As a volatility encoding of the meteorological forecast data; Step S34: and The decoder is input, and the two are first fused together for time... Calculate the fusion volatility code : in, , , Indicates a splicing operation; Step S35: Incorporate the volatility encoding After two floors Advanced meteorological characterization was obtained. , The calculation formula is as follows: 。 3. The method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks according to claim 2, characterized in that, Step S4: Inputting the advanced meteorological characterization into a time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day, specifically including: Advanced meteorological characterization Input a time-specific prediction network to calculate the predicted fluctuation sequence of photovoltaic power generation within the next day according to the following formula: in, express Advanced meteorological representation of time The first in OK; and These are specific parameters; For prediction The fluctuation value of photovoltaic power generation at any given time; For all moments of the day Each uses specific prediction parameters, so the predicted power fluctuation value for that day is... for: 。 4. The method for predicting photovoltaic power generation the following day based on seasonal decomposition and convolutional networks according to claim 3, characterized in that, Step S5: Add the predicted fluctuation sequence to the historically measured seasonal trend sequence to obtain the final photovoltaic power prediction result, specifically including: Step S51: From the historically measured seasonal trend sequence: Obtain the seasonal trend sequence of power generation corresponding to the power generation characteristics of the power plants. ; Step S52: ... and By summing the results at each time step, the final power prediction for the next day can be obtained. for: Step S53: Construct the L2 loss function as the loss function for photovoltaic power prediction. The calculation formula is as follows: in, for Predicted photovoltaic power at any time for Actual photovoltaic power at any given time.

5. A next-day photovoltaic power generation prediction system based on seasonal decomposition and convolutional networks, characterized in that, Includes the following modules: A module for constructing historical measured data and meteorological forecast data is established to collect historical measured meteorological data and power generation during photovoltaic power generation. This data is then combined with meteorological forecasts provided by the meteorological observatory and preprocessed to construct the historical measured data and meteorological forecast data, specifically including: Step S11: Collect the previous... The measured meteorological data and power generation during the photovoltaic power generation process within the day, wherein the measured meteorological data includes five characteristic values ​​measured every 15 minutes: irradiance, temperature, humidity, air pressure and wind speed, and the power generation includes one characteristic value every 15 minutes: power generation of the power station. Constructing the first Historical measured data of Japan ;in Indicates the previous number day Historical measured data at any given time ; Step S12: The historical measured data is processed using a normalization method to scale the numerical values, adjusting the range to [0,1], and constructing the historical measured data for the previous 15 days. ; Step S13: Collect photovoltaic weather forecast data for the next day, which includes five characteristic values ​​predicted every 15 minutes: irradiance, temperature, humidity, air pressure, and wind speed. Constructing next day's weather forecast data ,in Indicates the number of steps forward. day Predicted data for each moment, ; Step S14: The numerical scale of the next day's weather forecast data is scaled down using a normalization method, adjusting the numerical range to [0,1]; the forecast data for the previous 14 days and the weather forecast data for the day to be predicted are constructed. ,in This is the forecast data for the date to be predicted; A module for constructing historical measured fluctuation sequences and meteorological forecast fluctuation sequences is used to perform seasonal sequence decomposition on the historical measured data and the meteorological forecast data, respectively, to obtain the historical measured seasonal trend sequence and the historical measured fluctuation sequence, as well as the meteorological forecast fluctuation sequence. Specifically, this includes: Step S21: Calculate the historical measured data In each Mean at time: This yields a historically measured seasonal trend sequence: The latest historical measurement data is the observation data from one day prior. The historically measured fluctuation sequence is obtained by subtracting the historically measured seasonal trend sequence from this sequence: Step S22: Calculate the meteorological forecast data. In each Mean at time: The seasonal trend series of meteorological forecasts were obtained. The latest weather forecast data is based on observational data for the day to be forecasted. The fluctuation sequence of the weather forecast is obtained by subtracting the seasonal trend sequence of the weather forecast from this sequence: ； A high-level meteorological characterization module is used to: input the historical measured fluctuation sequence and the meteorological forecast fluctuation sequence into a multi-kernel gated convolutional neural network for encoding-decoding operations to obtain a high-level meteorological characterization; A module for obtaining photovoltaic power generation prediction fluctuation sequence is used to input the advanced meteorological characterization into a time-specific prediction network to calculate the prediction fluctuation sequence of photovoltaic power generation within the next day. The photovoltaic power prediction module is used to add the predicted fluctuation sequence to the historical measured seasonal trend sequence to obtain the final photovoltaic power prediction result.

Images