Photovoltaic power generation time sequence prediction method based on periodic modeling and channel interaction

Abstract

The invention discloses a photovoltaic power generation time sequence prediction method based on periodic modeling and channel interaction, and the method comprises the steps: collecting and processing the historical data of photovoltaic power generation; then, extracting a learnable periodic mode in the time sequence, and removing a periodic component to obtain a residual component; then, through reversible instance normalization, sequence embedding, a channel interaction module and a linear predictor, modeling and prediction of a residual component are completed; and finally, adding the predicted residual component and the periodic component to generate a final prediction result. Through an aggregated channel interaction strategy, the calculation complexity is reduced while the channel correlation is captured, the dependence on an abnormal channel is reduced, and the robustness and the expansion capability of the model are improved; meanwhile, by directly modeling a periodic mode in the time sequence, the capability of extracting inherent periodicity in the photovoltaic power generation time sequence data is improved, so that future power generation data is predicted more accurately; the method can be widely applied to optimal scheduling and management of a photovoltaic power generation system.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence and new energy, and particularly relates to a photovoltaic power generation time series prediction method based on periodic modeling and channel interaction. Background Art

[0002] Solar energy, as a renewable energy source, has been receiving increasing attention. Since the photovoltaic power generation is greatly affected by external temperature, wind speed, wind direction, as well as time and season changes, it has seasonal periodicity, time periodicity, and uncertainty. However, affected by factors such as weather and seasons, the changes in solar illumination and meteorological conditions will cause non-linear changes and fluctuations in the output power of photovoltaic power generation, and the power output is not stable enough. Without photovoltaic power generation prediction, it will lead to unstable power supply, difficult load scheduling, and increased risks in power grid operation. Therefore, photovoltaic power generation prediction is indispensable in power system management.

[0003] Although existing time series prediction methods have been widely used and proven effective in photovoltaic power generation prediction, there are still some problems: First, existing models usually emphasize their ability to extract long-term dependence features. Models such as Informer, Autoformer, and PatchTST utilize the advantages of Transformer in long-distance modeling to handle LTSF tasks. However, in order to extract their long-term dependence features, most of the structures are complex and the number of parameters is large; Second, nowadays, the research on photovoltaic power generation algorithms based on machine learning and deep learning algorithms all focus on emphasizing the advantages of channel independence against distribution drift, but ignore channel correlation, which limits further enhancement. Some methods use mechanisms such as attention or mixer to solve this problem by capturing channel correlation. Although these modules directly compare the characteristics of each pair of channels, they face quadratic complexity related to the number of channels. In addition, such a distributed structure may lack robustness in the presence of abnormal channels because they rely too much on the correlation between channels. Summary of the Invention

[0004] In order to solve the deficiencies of the prior art, achieve the purpose of efficiently extracting the inherent periodic patterns of data and enhancing the interaction between channels, thereby significantly improving the prediction accuracy of photovoltaic power generation and reducing the computational complexity, the present invention adopts the following technical solutions:

[0005] A photovoltaic power generation time series prediction method based on periodic modeling and channel interaction, comprising the following steps:

[0006] Step S1: Perform periodic pattern modeling on the time series data affecting photovoltaic power generation to obtain periodic patterns that show regular repeated changes within a certain time range;

[0007] Step S2: Subtract the learned periodic component from the obtained original time series data to obtain a residual component;

[0008] Step S3: Use an aggregated model of channel interaction to predict the residual components, obtaining a prediction result. Channel interaction is the modeling and utilization of the potential correlation relationships between different variables (channels). This kind of interaction modeling can capture the interactions between variables, improve the understanding of the overall data pattern, and thus enhance the prediction accuracy. Channel interaction aims to capture the correlation relationships between these variables; channel interaction is particularly important for time series modeling and prediction because the occurrence of many phenomena is the result of the combined action of multiple variables; directly ignoring the relationships between channels may lead to: 1) The prediction ability of the model decreases: The information interaction between variables cannot be fully utilized; 2) Insufficient robustness to abnormal data: Relying on single-channel prediction may be more vulnerable to noise; 3) Limited understanding of complex phenomena: For example, there may be non-linear or high-order interaction relationships between some variables, which must be captured through modeling; Since the distributed modeling method using the attention mechanism in the past has too high complexity and lacks robustness, the present invention proposes to use an aggregated model of channel interaction to predict the residual components, for capturing the relationships between environmental factors, improving the prediction ability in abnormal situations, and reducing the model calculation complexity; The specific steps of the aggregated modeling method are as follows:

[0009] Step S3.1: Normalization operation. In time series prediction, remove the historical local statistics to stabilize the prediction of the base predictor and restore these statistics to the prediction of the aggregated model.

[0010] Step S3.2: Sequence embedding.

[0011] Step S3.3: Channel interaction; Exchange information between channels through an aggregated star module and perform scheduling fusion with a single sequence to achieve the effect of channel interaction.

[0012] Step S3.4: Linear predictor, used to generate the predicted power of photovoltaic power generation.

[0013] Step S4: Add the predicted power back to the periodic component to obtain the finally predicted power generation.

[0014] Furthermore, in the step S1, explicitly model the periodic pattern directly; Given multiple (D) channels with a prior period length W, first generate a learnable cyclic period Q ∈ R W×D, all cycle periods are initialized to zero. These cycle periods are globally shared within the channel. By performing cycle replication, the cycle components C of the time series X of the same length are obtained. These cycle periods Q of length W are trained with the backbone module for time series prediction through gradient backpropagation, generating a learned representation (different from the initially initialized zero) that reveals the internal cycle pattern of the sequence; the cycle length W depends on the prior characteristics of the data and is the maximum stable period in the data. Considering that scenarios requiring long-term prediction usually exhibit prominent and distinct cycles, determining the specific cycle length is available and straightforward.

[0015] Furthermore, in the step S1, the period of the data is further examined through the autocorrelation function to measure the correlation between the time series and its lagged values, indicating the existence of autocorrelation within the data. The formula is as follows:

[0016]

[0017] where N represents the total number of observations, x t represents the value of the time series at time t, k represents the lag time, represents the mean value of the time series values. When the lag time k is consistent with the period of the data, the autocorrelation function value shows a significant peak, and the largest peak corresponds to the lag, and the lag is consistent with the length of the largest period existing in the data. Conversely, if the data lacks periodicity, no obvious peaks or valleys will appear.

[0018] Furthermore, in the step S1, data analysis is performed on the data affecting photovoltaic power generation, and data preprocessing is carried out according to the data analysis results, including replacing and filling outliers and missing values using linear interpolation; for time data, it is directly filled according to the real time; consistency checking is performed to find the data outside the reasonable range in the data and delete it.

[0019] Furthermore, the cycle pattern in the step S1 has the following characteristics:

[0020] Regularity: The cycle pattern is manifested as the data repeatedly showing similar fluctuation forms at certain time intervals. For photovoltaic power generation data, it shows daily periodicity due to day and night alternation;

[0021] Stability: If the source of periodicity is stable (such as the solar cycle in photovoltaic power generation), the cycle pattern does not change significantly over time;

[0022] Multiple periodicity: Some data has multiple periodicities superimposed. Photovoltaic power generation data contains daily cycles, weekly cycles, and seasonal cycles.

[0023] In long-term time series prediction, the periodic pattern is the core information because it reflects the long-term change law of the data. By extracting the periodic pattern, the data trend of the next period can be accurately predicted; the periodic pattern is usually relatively stable and can help the model ignore short-term anomalies; accurate modeling of the periodic pattern enables the model to better understand the repetitive law of the data, thereby improving the prediction accuracy.

[0024] Furthermore, in step S2, the original time series x is input t-L+1:t Remove the circulating component c t-L+1:t , and get the residual component x′ t-L+1:t ; The cyclic component C is a virtual sequence derived from the cyclic replication of the cyclic period Q. Since the above subsequence c cannot be directly obtained t-L+1:t Therefore, the cycle period Q needs to be aligned and repeated to obtain the cycle component c t-L+1:t The equivalent subsequence of includes the following steps:

[0025] Step S2.1: Shift the cycle period Q to the left by t mod W to obtain Q (t) , t represents time, t mod W can be regarded as the relative position index of the current sequence sample in Q;

[0026] Step S2.2: Repeat Q (t) Operate [L / W] times, and connect in series These two equivalent subsequences can be expressed as:

[0027]

[0028] Among them, W represents the cycle length, H represents the predicted time window length, and L represents the input time series length.

[0029] In step S4, the periodic alignment and repetition after adding back the periodic component are consistent with step S2.

[0030] Furthermore, in step S3.1, reversible instance normalization is used to calculate the residual component x′ obtained in step S2. t-L+1:t , center the series around zero mean, scale them to have unit variance, and back-normalize them on the forecast series.

[0031] Furthermore, the sequence embedding in step S3.2 is equivalent to setting the block length to the length of the entire sequence. Unlike block embedding, sequence embedding does not generate additional dimensions, so it is less complex than block embedding. In the present invention, sequence embedding is performed on the lookback window, and the sequence of each channel is embedded into S0=R using linear projection. C×d , where d is the hidden dimension:

[0032] S0 = Embedding(X)

[0033] Among them, S0 represents the result of sequence embedding, and X represents the time series of each channel.

[0034] Furthermore, in the step S3.3, given a multivariate sequence {S1, S2,.., S n}, its core representation o is a vector generated by an arbitrary function f of the following form:

[0035] o = f(S1, S2,.., S n )

[0036] The core representation o encodes the global information of all channels:

[0037] o i = Stoch_Pool(MLP1(S i-1 ))

[0038] Among them, MLP1: R d → R d′ is a projection that projects the sequence representation from the sequence hidden dimension d to the core dimension d′, consisting of two layers of the hidden dimension d and GELU activation. R represents the set of matrices of the sequence, MLP1 represents the first multi-layer perceptron, and Stoch_Pool represents the stochastic pooling operation. By aggregating the representations of n sequences, the core representation o ∈ R d′ is obtained. The stochastic pooling combines the advantages of mean pooling and max pooling;

[0039] Fuse the core and the representations of all sequences:

[0040] F i = Repeat_Concat(S i-1 , o i )

[0041] S i = MLP2(F i ) + S i-1

[0042] Among them, the Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain the feature F i ∈ R C ×(d+d′) , and then uses the second multi-layer perceptron MLP2 to fuse the concatenated representations to obtain S i ∈ R C×d , and the second multi-layer perceptron MLP2: R d+d′ → R dThe sequence representation is projected back from the concatenated sequence hidden dimension d + the core dimension d' to the sequence hidden dimension d. Like many deep learning modules, the present invention also adds a residual connection from the input to the output.

[0043] Further, in step S3.4, after channel interaction, a linear predictor is used to generate the predicted power of photovoltaic power generation

[0044]

[0045] where Linear represents a linear mapping function, and S N represents the output sequence of channel interaction of the Nth layer.

[0046] The advantages and beneficial effects of the present invention are as follows:

[0047] The present invention predicts future data based on historical time-series data information, performs explicit modeling on the periodic patterns in photovoltaic power generation time-series data, and then subtracts the data from the modeled periodic pattern components to obtain residual components, thereby avoiding the complex problem of periodic pattern extraction while improving efficiency and accuracy; then, an aggregated channel interaction method is used for the residual components to predict intermediate results, and finally the intermediate results are added back to the periodic pattern components to obtain the final prediction result. Compared with the traditional method of using distributed methods to process channel interaction information, the present invention can not only avoid the quadratic complexity related to the number of channels, but also overcome the problem that the distributed structure may lack robustness in the presence of abnormal channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flowchart of the method of an embodiment of the present invention.

[0049] Figure 2 is a schematic structural diagram of the prediction model constructed in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The following further details the specific embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0051] When dealing with the internal periodic patterns of data, existing time series photovoltaic power generation prediction methods usually emphasize the ability to extract long-term dependence features, and obtain periodic information through these long-term dependence features. However, they often have complex structures and require a large number of parameters. For example, the PatchTST model uses a Transformer to achieve prediction. Therefore, the present invention proposes a photovoltaic power generation time series prediction method based on periodic modeling and channel interaction, which directly performs explicit modeling on the periodic patterns in the data, uses learnable cyclic periods to explicitly model the inherent periodic patterns in the time series data, and then predicts its residual part, which can avoid the problem of complex periodic pattern extraction while improving efficiency and accuracy. At the same time, in order to consider the correlation between channels, an aggregated interaction model is used to predict the residual part, that is, a core is used to interact and process the information of each channel. Compared with the existing technology that uses a distributed method to process channel interaction information, it can not only avoid the quadratic complexity related to the number of channels, but also overcome the problem that the distributed structure may lack robustness in the presence of abnormal channels. As Figure 1 、 Figure 2 shown, the photovoltaic power generation time series prediction method of the present invention specifically includes the following steps:

[0052] Step S1: Preprocess the data affecting photovoltaic power generation, and perform periodic pattern modeling on the processed data to obtain its learnable periodic pattern.

[0053] First, analyze the data set, including data distribution, outliers, missing value situations, etc. The collected data contains a large amount of unreasonable data, including outlier processing and missing value processing. Linear interpolation can be used for replacement and filling. For time data, it can be directly filled according to the real time; consistency check, find the data that is not within the reasonable range in the original data, and for the convenience of algorithm training, it can be directly deleted.

[0054] The periodic pattern in time series data refers to the regular repeated change pattern presented by the data within a certain time range. This regularity may stem from natural phenomena, seasonal changes or human activities. It has the following characteristics:

[0055] Regularity: The periodic pattern is manifested as the data repeatedly showing similar fluctuation forms at certain time intervals. Photovoltaic power generation data usually shows daily periodicity due to day and night alternation.

[0056] Stability: If the source of periodicity is relatively stable (such as the solar cycle in photovoltaic power generation), the periodic pattern usually does not change significantly over time.

[0057] Multiple periodicity: Some data may have multiple periodicities superimposed. Photovoltaic power generation data may contain daily, weekly and seasonal periods.

[0058] In long-term time series prediction, the periodic pattern is the core information because it reflects the long-term variation law of the data. By extracting the periodic pattern, the data trend of the next period can be accurately predicted. The periodic pattern is usually relatively stable, which can help the model ignore short-term anomalies. The accurate modeling of the periodic pattern enables the model to better understand the repetitive pattern of the data, thereby improving the prediction accuracy. For this reason, we propose a method for explicitly modeling the periodic pattern directly.

[0059] Given D channels with a prior cycle length W, first generate a learnable cyclic period Q ∈ R W×D , and all cyclic periods are initialized to zero. These cyclic periods are globally shared within the channels, which means that by performing cyclic replication, the cyclic component C of the sequence X of the same length can be obtained. These cyclic periods Q of length W are trained with gradient backpropagation together with the backbone module for prediction, generating a learned representation (different from the initially initialized zero) that reveals the internal cyclic pattern of the sequence.

[0060] The cycle length W depends on the prior characteristics of the dataset and is set to the maximum stable cycle in the dataset. Considering that scenarios requiring long-term prediction usually exhibit prominent and distinct cycles (e.g., solar radiation intensity, ambient temperature, weather conditions), it is available and straightforward to determine the specific cycle length. In addition, the cycle of the dataset can be further examined through the autocorrelation function (ACF).

[0061] The autocorrelation function (ACF) is a powerful mathematical tool that can help us determine the periodicity within the data. The autocorrelation function measures the correlation between a time series and its lagged values, indicating the existence of autocorrelation within the data. Mathematically, this can be expressed as:

[0062]

[0063] where N represents the total number of observations, x t represents the value of the time series at time t, k represents the lag time, represents the mean value of the time series values.

[0064] When the lag time k is consistent with the period of the data, the ACF value shows a significant peak. Specifically, the largest peak corresponds to the lag, and the lag is consistent with the length of the largest period existing in the dataset. Conversely, if the data lacks periodicity, no obvious peaks or valleys will appear.

[0065] Step S2: Subtract the learned cyclic component from the original input data to obtain the residual component.

[0066] From the original input x t-L+1:tRemove the cyclic component c t-L+1:t to obtain the residual component x' t-L+1:t , where t represents time and L represents the length of the input time series.

[0067] The cyclic component C is a virtual sequence derived from the cyclic replication of Q. Since the above subsequence c cannot be directly obtained t-L+1:t . Therefore, it is necessary to appropriately align and repeat the cycle period Q to obtain an equivalent subsequence, which specifically includes the following steps:

[0068] Step S2.1: Obtain Q by shifting Q to the left by the position of t mod W (t) ; t mod W can be regarded as the relative position index of the current sequence sample within Q

[0069] Step S2.2: Repeat Q (t) operation [L / W] times and concatenate These two equivalent subsequences can be expressed as:

[0070]

[0071]

[0072] where H represents the length of the predicted time window.

[0073] Step S3: Use the aggregated model of channel interaction to predict the residual component to obtain the prediction result.

[0074] Channel Interaction refers to the modeling and utilization of the potential correlation relationships between different variables (channels). This interaction modeling can capture the interactions between variables, improve the understanding of the overall data pattern, and thus enhance the prediction accuracy. Channel Interaction aims to capture the correlation relationships between these variables. For example, in photovoltaic power generation prediction, an increase in temperature may increase the contribution of solar radiation to power generation, and the cloud thickness and solar radiation intensity jointly determine the power generation efficiency.

[0075] Channel Interaction is particularly important for time series modeling and prediction because the occurrence of many phenomena is the result of the combined action of multiple variables. Directly ignoring the relationships between channels may lead to: 1) a decline in the prediction ability of the model: unable to fully utilize the information interaction between variables; 2) insufficient robustness to abnormal data: relying on single-channel prediction may be more vulnerable to noise; 3) limited understanding of complex phenomena: for example, there may be non-linear or high-order interaction relationships between some variables, which must be captured through modeling.

[0076] Due to the high complexity and lack of robustness of previous distributed modeling methods using attention mechanisms, the present invention proposes an aggregative modeling method, whose functions include: 1) capturing the relationships between environmental factors: such as the interaction between cloud thickness and solar radiation, the influence of temperature on wind speed, and the comprehensive influence on the overall power generation efficiency; 2) improving the prediction ability in abnormal situations: when the data of a certain channel is abnormal, the model can still use other channels for accurate prediction; 3) reducing the computational complexity of the model: through aggregative modeling, the complex relationships of all channel pairs are avoided while retaining key interaction information. The aggregative modeling method specifically includes the following steps:

[0077] Step S3.1: Invertible instance normalization: Normalization is a commonly used technique for calibrating the input data distribution. In time series prediction, historical local statistics are usually removed to stabilize the prediction of the base predictor and restore these statistics to the model prediction. Following the common practice of many advanced models, the present invention applies invertible instance normalization to the residual component x' obtained in step S2 t-L+1:t , center the sequence around zero mean, scale it to unit variance, and reverse normalize on the prediction sequence.

[0078] Step S3.2: Sequence embedding: Sequence embedding is an extreme case of the ubiquitous block embedding in time series, equivalent to setting the block length to the length of the entire sequence. Different from block embedding, sequence embedding does not generate additional dimensions and thus has lower complexity than block embedding. In the present invention, sequence embedding is performed on the backtracking window. Specifically, the present invention uses linear projection to embed the sequence of each channel into S0 = R C×d , where d is the hidden dimension:

[0079] S0 = Embedding(X)

[0080] where S0 represents the result of sequence embedding and X represents the sequence of each channel.

[0081] Step S3.3: Channel interaction: Information between channels is exchanged through an aggregative star module and scheduled and fused with a single sequence to achieve the effect of channel interaction.

[0082] Given a multivariate sequence {S1, S2,.., S n} with n channels, its core representation o is a vector generated by any function f of the following form:

[0083] o = f(S1, S2,.., S n )

[0084] The core representation o encodes the global information of all channels. To obtain such a representation, the present invention adopts the following form:

[0085] o i = Stoch_Pool(MLP1(S i-1 ))

[0086] where MLP1: R d → R d′ is a projection that projects the sequence representation from the sequence hidden dimension d to the core dimension d′, consisting of two layers of the hidden dimension d and the GELU activation. R represents the set of matrices of the sequence, and MLP represents the multi-layer perceptron. Stoch_Pool represents the stochastic pooling operation, which obtains the core representation o ∈ R d′ by aggregating the representations of n sequences. The stochastic pooling combines the advantages of mean pooling and max pooling. Next, the present invention fuses the core and the representations of all sequences:

[0087] F i = Repeat_Concat(S i-1 , o i )

[0088] S i = MLP2(F i ) + S i-1

[0089] where the Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain F i ∈ R C×(d+d′) . Then, another MLP (MLP2: R d+d′ → R d ) is used to fuse the concatenated representations and project them back to the hidden dimension d, i.e., S i ∈ R C×d . Like many deep learning modules, the present invention also adds a residual connection from the input to the output.

[0090] Step S3.4: Linear predictor: After channel interaction, a linear predictor is used to generate the prediction result. Assuming that the output sequence representation of the Nth layer is S N , then the prediction is calculated as:

[0091]

[0092] where represents the prediction result, i.e., the predicted power in the photovoltaic power generation prediction, and Linear represents the linear mapping function.

[0093] Step S4: Add the prediction result back to the periodic component to obtain the final prediction result, i.e., the generated power.

[0094] The cycle alignment and repetition after adding back the periodic component are the same as in step S2.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic power generation time series prediction method based on periodic modeling and channel interaction, characterized in that It includes the following steps: Step S1: Perform periodic pattern modeling on the time series data affecting photovoltaic power generation to obtain a periodic pattern that shows regular repeated changes within a certain time range; Step S2: Subtract the learned periodic component from the obtained original time series data to obtain a residual component; Step S3: Use an aggregated model of channel interaction to predict the residual component to obtain a prediction result. The specific steps of the aggregated modeling method are as follows: Step S3.1: Normalization operation, removing historical local statistics and restoring these statistics to the prediction of the aggregated model; Step S3.2: Sequence embedding; Step S3.3: Channel interaction; exchange information between channels through an aggregated module and perform scheduling fusion with a single sequence; Step S3.4: Linear predictor, used to generate the predicted power of photovoltaic power generation; Step S4: Add the predicted power back to the periodic component to obtain the finally predicted power generation power.

2. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, wherein: In step S1, an explicit model of the periodic pattern is directly constructed; given multiple channels with prior periodic lengths, learnable cyclic periods are first generated, which are globally shared within the channels. By performing cyclic replication, the cyclic components of time series with the same length are obtained. The cyclic periods and the backbone module used for time series prediction are trained together through gradient backpropagation to generate a learned representation that reveals the internal cyclic pattern of the sequence; the periodic length depends on the prior characteristics of the data and is the maximum stable period in the data.

3. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: In step S1, the period of the data is further examined through the autocorrelation function, which measures the correlation between a time series and its lagged values. The formula is as follows: where N represents the total number of observations, x t represents the value of the time series at time t, k represents the lag time, represents the mean value of the time series values. When the lag time k is consistent with the period of the data, the autocorrelation function value shows a peak, and the largest peak corresponds to the lag, and the lag is consistent with the length of the largest period existing in the data.

4. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: In step S1, data analysis is performed on the data affecting photovoltaic power generation, and data preprocessing is carried out according to the data analysis results, including replacing and filling outliers and missing values using linear interpolation; for time data, it is directly filled according to the real time; Consistency check, finding out the data that is not within the reasonable range in the data and deleting it.

5. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, wherein: The periodic pattern in step S1 has the following characteristics: Regularity: The periodic pattern shows that the data repeatedly exhibits similar fluctuation patterns at certain time intervals; Stability: If the periodic source is stable, the periodic pattern does not change significantly over time; Multiple periodicity: Some data has multiple periodic superpositions.

6. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, wherein: In the step S2, the cyclic component c is removed from the input original time series t-L+1:t , and a residual component is obtained; the cyclic component is derived from the cyclic replication of the cyclic period Q, and the cyclic period Q is aligned and repeated to obtain an equivalent subsequence of the cyclic component c t-L+1:t , including the following steps: Step S2.1: Obtain Q by shifting the cycle Q to the left at the position of t mod W, where t represents time; (t) , where t represents time; Step S2.2: Repeat Q (t) for [L / W] operations, and concatenate These two equivalent subsequences can be expressed as: Among them, W represents the period length, H represents the length of the prediction time window, and L represents the length of the input time series.

7. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: In step S3.1, reversible instance normalization is adopted. For the residual component obtained in step S2, the sequence is centered around zero mean, scaled to unit variance, and reverse-normalized on the prediction sequence.

8. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: The sequence embedding in step S3.2 is performed on the backtracking window, and linear projection is used to embed the sequence of each channel into S0: S0 = Embedding(X) where S0 represents the result of sequence embedding, and X represents the time series of each channel.

9. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, wherein: In the step S3.3, given a multivariate sequence {S1, S2,.., S n} with n channels, the core representation o is a vector generated by a function f of any of the following forms: o = f(S1, S2,.., S n ) The core representation o encodes the global information of all channels: o i = Stoch_Pool(MLP1(S i-1 )) Among them, MLP1:R d →R d′ is a projection that projects the sequence representation from the sequence hidden dimension d to the core dimension d′, R represents the set of matrices of the sequence, MLP1 represents the first multi-layer perceptron, Stoch_Pool represents the stochastic pooling operation, and the core representation o∈R is obtained by aggregating the representations of n sequences d′ ; Fuse the core and the representations of all sequences: F i = Repeat_Concat(S i-1 , o i ) S i = MLP2(F i ) + S i-1 Among them, the Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain the feature F i , and then uses the second multi-layer perceptron MLP2 to fuse the concatenated representations to obtain S i , the second multi-layer perceptron MLP2: R d+d′ →R d projects the sequence representation from the concatenated sequence hidden dimension d + core dimension d' back to the sequence hidden dimension d 10. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 9, characterized in that: In step S3.4, after channel interaction, a first-order linear predictor is used to generate the predicted power of photovoltaic power generation Among them, Linear represents a linear mapping function, and S N represents the channel interaction output sequence of the Nth layer.

Images