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

Abstract

The application discloses a photovoltaic power generation time series prediction method based on periodic modeling and channel interaction. Firstly, historical data of photovoltaic power generation is collected and processed. Then, learnable periodic patterns in the time series are extracted, and periodic components are removed to obtain residual components. Then, reversible instance normalization, sequence embedding, channel interaction module and linear predictor are used to complete modeling and prediction of the residual components. Finally, the predicted residual components and the periodic components are added to generate the final prediction result. Through the aggregated channel interaction strategy, the application can capture channel correlation while reducing computational complexity, reduce dependence on abnormal channels, and improve the robustness and expansion capability of the model. Meanwhile, by directly modeling the periodic patterns in the time series, the application improves the extraction capability of inherent periodicity in photovoltaic power generation time series data, thereby more accurately predicting future power generation data. The application can be widely applied to optimization scheduling and management of photovoltaic power generation systems.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and new energy, specifically involving a photovoltaic power generation time series prediction method based on periodic modeling and channel interaction. Background Technology

[0002] Solar energy is gaining increasing attention as a renewable energy source. However, photovoltaic (PV) power generation is significantly affected by external factors such as temperature, wind speed, wind direction, time, and seasonal variations, exhibiting seasonal and temporal periodicity and uncertainty. Furthermore, changes in solar irradiance and meteorological conditions, influenced by weather and seasons, cause nonlinear variations and fluctuations in PV power output, resulting in unstable power output. Without PV power generation forecasting, power supply instability can occur, load dispatching difficulties can arise, and grid operation risks can increase. Therefore, PV power generation forecasting is indispensable for power system management.

[0003] While existing time series forecasting methods have been widely used and proven effective in photovoltaic (PV) power generation forecasting, several issues remain: First, existing models typically emphasize their ability to extract long-term dependency features. Models like Informer, Autoformer, and PatchTST leverage the Transformer's strength in long-range modeling to handle LTSF tasks. However, to extract these long-term dependency features, most are structurally complex and have a large number of parameters. Second, current research on PV power generation algorithms based on machine learning and deep learning focuses on the advantage of channel independence against distribution drift, but neglects channel correlation, limiting further enhancements. Some methods address this issue by capturing channel correlation using mechanisms such as attention or mixers. While these modules directly compare the characteristics of each pair of channels, they face quadratic complexity related to the number of channels. Furthermore, such distributed structures may lack robustness in the presence of anomalous channels because they rely too heavily on the correlation between channels. Summary of the Invention

[0004] To address the shortcomings of existing technologies, achieve efficient extraction of inherent periodic patterns from data, and enhance inter-channel interaction, thereby significantly improving the prediction accuracy of photovoltaic power generation and reducing computational complexity, this invention adopts the following technical solution:

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

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

[0007] Step S2: Subtract the learned periodic component from the acquired raw time series data to obtain the residual component;

[0008] Step S3: Predict the residual components using a convergent model of channel interaction to obtain the prediction results. Channel interaction is the modeling and utilization of potential correlations between different variables (channels). This interactive modeling can capture the interactions between variables, improve the understanding of the overall data pattern, and thus improve prediction accuracy. Channel interaction aims to capture the correlations between these variables. Channel interaction is particularly important for time series modeling and prediction because many phenomena are the result of the combined effects of multiple variables. Directly ignoring the relationships between channels may lead to: 1) Decreased predictive power of the model: failure to fully utilize the information interaction between variables; 2) Insufficient robustness to anomalous data: reliance on a single channel for prediction may be more susceptible to noise; 3) Limited understanding of complex phenomena: for example, there may be nonlinear or higher-order interaction relationships between some variables, which must be captured through modeling. Since previous distributed modeling methods using attention mechanisms are too complex and lack robustness, this invention proposes to use a convergent model of channel interaction to predict the residual components, in order to capture the relationships between environmental factors, improve the predictive power under anomalous conditions, and reduce the computational complexity of the model. The convergent modeling method specifically includes the following steps:

[0009] Step S3.1: Normalization operation, in time series forecasting, removes historical local statistics to stabilize the forecasts of the base forecasters and restores these statistics to the forecasts of the aggregate model;

[0010] Step S3.2: Sequence embedding;

[0011] Step S3.3: Channel interaction; Information between channels is exchanged through an aggregated star module and scheduled and fused with a single sequence to achieve channel interaction;

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

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

[0014] Furthermore, in step S1, the periodic pattern is explicitly modeled directly; given multiple (D) channels with prior period length W, a learnable cyclic period Q∈R is first generated. W×DAll cyclic periods are initialized to zero and are globally shared within the channel. Cyclic components C of the same length for the time series X are obtained by performing cyclic replication. These cyclic periods Q of length W are trained together with the backbone module used for time series prediction using gradient backpropagation, producing a learned representation that reveals the cyclical patterns within the sequence (different from the initially initialized zero). The period length W depends on the prior characteristics of the data and is the largest stable period in the data. Considering that scenarios requiring long-term prediction often exhibit prominent and well-defined periods, determining the specific period length is both feasible and straightforward.

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

[0016]

[0017] Where N represents the total number of observations, x t This represents the value of the time series at time t, where k represents the lag time. The autocorrelation function represents the mean 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. The largest peak corresponds to the lag, which is consistent with the length of the largest period in the data. Conversely, if the data lacks periodicity, there will be no obvious peaks or troughs.

[0018] Furthermore, in step S1, data analysis is performed on the data affecting photovoltaic power generation, and data preprocessing is performed based on the data analysis results, including replacing and filling out outliers and missing values ​​using linear interpolation; for time data, it is directly filled according to the actual time; consistency check is performed to find data that is not within a reasonable range and delete it.

[0019] Furthermore, the periodic pattern in step S1 includes the following characteristics:

[0020] Regularity: Periodic patterns are characterized by data repeatedly exhibiting similar fluctuations within certain time intervals. For photovoltaic power generation data, due to the alternation of day and night, it shows a daily periodicity.

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

[0022] Multiple periodicity: Some data exhibit multiple overlapping periodicity patterns, including daily, weekly, and seasonal periods in photovoltaic power generation data.

[0023] In long-term time series forecasting, cyclical patterns are core information because they reflect the long-term changing patterns of the data. By extracting cyclical patterns, the data trend in the next period can be accurately predicted. Cyclical patterns are usually relatively stable and can help the model ignore short-term anomalies. Accurate modeling of cyclical patterns enables the model to better understand the recurring patterns of the data, thereby improving the prediction accuracy.

[0024] Further, in step S2, from the input original time series x t-L+1:t Remove circulating component c t-L+1:t The residual component x′ is obtained. t-L+1:t The cyclic component C is a virtual sequence derived from the cyclic replication of the cyclic period Q. Since the aforementioned subsequence c cannot be obtained directly... t-L+1:t Therefore, it is necessary to align and repeat the cycle period Q to obtain the cycle component c. t-L+1:t The equivalent subsequences are obtained by the following steps:

[0025] Step S2.1: Obtain Q by shifting the position to the left by the cycle period Q using t mod W. (t) t represents time, and 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) Operation [L / W] times, and connected in series These two equivalent subsequences can be represented as:

[0027]

[0028] Where W represents the period length, H represents the prediction time window length, and L represents the input time series length.

[0029] In step S4, the period alignment and repetition after adding back the periodic components are the same as in step S2.

[0030] Furthermore, in step S3.1, reversible instance normalization is used for the residual component x′ obtained in step S2. t-L+1:t The sequence is centered at zero mean, scaled to unit variance, and then inversely normalized on the predicted sequence.

[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, thus having lower complexity. In this invention, sequence embedding is performed on a backtracking window, using linear projection to embed the sequence of each channel into S0 = R. C×d In the context of , where d is the hidden dimension:

[0032] S0 = Embedding(X)

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

[0034] Further, in step S3.3, a multivariate sequence {S1, S2, ..., S...} with n channels is given. n}, its core representation is that o is a vector generated by any function f with the following form:

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

[0036] The core representation 'o' encodes global information across all channels:

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

[0038] Among them, MLP1:R d →R d′ It projects the sequence representation from the hidden dimension d to the core dimension d′, consisting of two layers: the hidden dimension d and GELU activation. R represents the set of sequence matrices, MLP1 represents the first multilayer perceptron, and Stoch_Pool represents the random pooling operation. The core representation o∈R is obtained by aggregating the representations of n sequences. d′ Random pooling combines the advantages of mean pooling and max pooling.

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

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

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

[0042] The Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain the feature F. i ∈R C ×(d+d′) Then, a second multilayer perceptron (MLP2) is used to fuse the concatenated representations to obtain S. i ∈R C×d The second multilayer perceptron MLP2:R d+d′ →R dProjecting the sequence representation back from the concatenated sequence hidden dimension d + core dimension d′ back to the sequence hidden dimension d, and like many deep learning modules, this invention also adds a residual connection from input to output.

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

[0044]

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

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

[0047] This invention predicts future data based on historical time-series data. It explicitly models the periodic patterns in photovoltaic power generation time-series data, then subtracts the modeled periodic pattern components from the data to obtain residual components, thereby avoiding the complexities of periodic pattern extraction while improving efficiency and accuracy. Next, an aggregated channel interaction method is used to predict intermediate results on the residual components. Finally, the intermediate results are added back to the periodic pattern components to obtain the final prediction result. Compared to traditional distributed methods for processing channel interaction information, this invention not only avoids the secondary complexity related to the number of channels but also overcomes the potential lack of robustness of distributed structures in the presence of abnormal channels. Attached Figure Description

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

[0049] Figure 2 This is a schematic diagram of the structure of the prediction model constructed in an embodiment of the present invention. Detailed Implementation

[0050] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] Existing time-series photovoltaic (PV) power generation forecasting methods typically emphasize the ability to extract long-term dependent features when processing inherent periodic patterns in the data. These features are used to obtain periodic information, but they often involve complex structures and require a large number of parameters. For example, the PatchTST model uses a Transformer for prediction. Therefore, this invention proposes a PV power generation time-series forecasting method based on periodic modeling and channel interaction. This method directly and explicitly models the periodic patterns in the data, using learnable cyclic periods to explicitly model the inherent periodic patterns in the time-series data, and then predicts the residuals. This avoids the complexity of periodic pattern extraction while improving efficiency and accuracy. Furthermore, to consider the correlation between channels, an aggregated interaction model is used to predict the residuals. This means a single core is used to interact and process information from various channels. Compared to existing technologies that use distributed methods to process channel interaction information, this not only avoids the secondary complexity related to the number of channels but also overcomes the potential lack of robustness of distributed structures in the presence of abnormal channels. Figure 1 , Figure 2 As 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 model the processed data into a periodic pattern to obtain its learnable periodic pattern.

[0053] First, the dataset is analyzed, including data distribution, outliers, and missing values. The collected data contains a large amount of unreasonable data. Outlier and missing value handling can be performed using linear interpolation for replacement and imputation. For time-based data, it can be directly imputed to reflect the actual time. Consistency checks are conducted to identify data outside the reasonable range in the original dataset; these can be deleted to facilitate algorithm training.

[0054] Periodic patterns in time series data refer to recurring patterns of change within a certain time frame. These patterns may originate from natural phenomena, seasonal variations, or human activities. They have the following characteristics:

[0055] Regularity: Periodic patterns are characterized by data exhibiting similar fluctuations repeatedly within certain time intervals. Photovoltaic power generation data typically shows a daily periodicity due to the day-night cycle.

[0056] Stability: If the source of the 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 periodicities: Some data may exhibit multiple overlapping periodicities. Photovoltaic power generation data may include daily, weekly, and seasonal periods.

[0058] In long-term time series forecasting, cyclical patterns are core information because they reflect the long-term variation patterns of the data. By extracting cyclical patterns, the data trend of the next period can be accurately predicted. Cyclical patterns are usually relatively stable and can help the model ignore short-term anomalies. Accurate modeling of cyclical patterns enables the model to better understand the recurring patterns of the data, thereby improving the prediction accuracy. To this end, we propose a method for directly and explicitly modeling cyclical patterns.

[0059] Given D channels with prior period length W, first generate learnable cyclic periods Q∈R. W×D All cyclic periods are initialized to zero. These cyclic periods are globally shared within the channel, meaning that by performing cyclic replication, cyclic components C of the same length of sequence X can be obtained. These cyclic periods Q of length W are trained together with the backbone module used for prediction through gradient backpropagation, producing a learned representation that reveals the cyclic patterns within the sequence (different from the initially initialized zero).

[0060] The period length W depends on the prior characteristics of the dataset and is set to the maximum stable period in the dataset. Considering that scenarios requiring long-term predictions typically exhibit prominent, well-defined periods (e.g., solar radiation intensity, ambient temperature, weather conditions), determining a specific period length is readily available and straightforward. Furthermore, the period of the dataset can be further examined using the autocorrelation function (ACF).

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

[0062]

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

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

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

[0066] From the original input x t-L+1:tRemove circulating component c t-L+1:t The residual component x′ is obtained. t-L+1:t 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 copying of Q. Since the aforementioned subsequence C cannot be obtained directly... t-L+1:t Therefore, it is necessary to properly align and repeat the cyclic period Q to obtain equivalent subsequences, specifically including the following steps:

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

[0069] Step S2.2: Repeat Q (t) Operation [L / W] times, and connected in series These two equivalent subsequences can be represented as:

[0070]

[0071]

[0072] Where H represents the length of the prediction time window.

[0073] Step S3: Use the aggregation model of channel interaction to predict the residual components and obtain the prediction results.

[0074] Channel interaction refers to the modeling and utilization of potential correlations between different variables (channels). This interactive modeling can capture the interactions between variables, improve the understanding of overall data patterns, and thus enhance prediction accuracy. Channel interaction aims to capture the correlations between these variables. For example, in photovoltaic power generation prediction, rising temperatures may increase the contribution of solar radiation to power generation, and cloud thickness and solar radiation intensity together determine power generation efficiency.

[0075] Channel interactions are particularly important for time series modeling and prediction because many phenomena are the result of the combined effects of multiple variables. Directly ignoring the relationships between channels may lead to: 1) decreased predictive power of the model: failure to fully utilize the information interaction between variables; 2) insufficient robustness to anomalous data: reliance on a single channel for prediction may be more susceptible to noise; 3) limited understanding of complex phenomena: for example, there may be nonlinear or higher-order interactions between some variables, which must be captured through modeling.

[0076] Because previous distributed modeling methods using attention mechanisms were too complex and lacked robustness, this invention proposes an aggregated modeling method. Its 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 their combined impact on overall power generation efficiency; 2) improving predictive capabilities under abnormal conditions: when data from one channel is abnormal, the model can still make accurate predictions using other channels; 3) reducing model computational complexity: through aggregated modeling, the complex relationships between all channel pairs are avoided, while retaining key interaction information. The aggregated modeling method specifically includes the following steps:

[0077] Step S3.1: Reversible Instance Normalization: Normalization is a common technique for calibrating the distribution of input data. In time series forecasting, historical local statistics are typically removed to stabilize the forecasts of base forecasters and then restored to the model predictions. Following the common practice of many advanced models, this invention applies reversible instance normalization for the residual component x′ obtained in step S2. t-L+1:t The sequence is centered at zero mean, scaled to unit variance, and then inversely normalized on the predicted sequence.

[0078] Step S3.2: Sequence Embedding: Sequence embedding is an extreme case of block embedding, which is common in time series analysis. It is equivalent to setting the block length to the length of the entire sequence. Unlike block embedding, sequence embedding does not introduce additional dimensions, thus having lower complexity. In this invention, sequence embedding is performed on a backtracking window. Specifically, this invention uses linear projection to embed the sequence of each channel into S0 = R. C×d In the context of , 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 aggregated star topology module and scheduled and merged with a single sequence to achieve the function of channel interaction.

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

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

[0084] The core representation 'o' encodes global information across all channels. To obtain this representation, the present invention employs the following form:

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

[0086] Among them, MLP1:R d →R d′ It projects the sequence representation from the hidden dimension d to the core dimension d′, consisting of two layers: the hidden dimension d and GELU activation. R represents the set of matrices of the sequence, and MLP represents a multilayer perceptron. Stoch_Pool represents the random pooling operation, which obtains the core representation o∈R by aggregating the representations of n sequences. d′ Random pooling combines the advantages of mean pooling and max pooling. Next, this invention fuses the representations of the core and all sequences:

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

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

[0089] The Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain F. i ∈R C×(d+d′) Then use another MLP (MLP2:R) d+d′ →R d The concatenated representations are then fused and projected back to the hidden dimension d, i.e., S. i ∈R C×d Like many deep learning modules, this invention also adds residual connections from input to output.

[0090] Step S3.4: Linear Predictor: After channel interaction, a linear predictor is used to generate the prediction result. Assume the output sequence of the Nth layer is represented as S. N Then predict The calculation is as follows:

[0091]

[0092] in, This represents the prediction result, i.e., the predicted power in photovoltaic power generation prediction. 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 power generation.

[0094] The period alignment and repetition after adding back the periodic components are consistent with step S2.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to 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... Includes the following steps: Step S1: Perform periodic pattern modeling on the time series data affecting photovoltaic power generation to obtain periodic patterns that exhibit regular repetitive changes within a certain time range; Step S2: Subtract the learned periodic component from the acquired raw time series data to obtain the residual component; Step S3: Predict the residual components using a convergent model of channel interaction to obtain the prediction results. The convergent modeling method specifically includes the following steps: Step S3.1: Normalization operation, remove historical local statistics and restore these statistics to the predictions of the aggregate model; Step S3.2: Sequence embedding; Perform sequence embedding on the backtracking window, using linear projection 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; Step S3.3: Channel interaction; information is exchanged between channels through aggregated modules and fused with a single sequence; specifically, given a multivariate sequence {S1, S2, ..., S...} with n channels... n }, its core representation is that o is a vector generated by any function f with the following form: o=f(S1,S2,..,S n ) The core representation 'o' encodes global information across all channels: O i <Stoch_Pool(MLP1(S i-1 )) Among them, MLP1:R d →R d′ It projects the sequence representation from the hidden dimension d to the core dimension d′, where R represents the set of matrices of the sequence, MLP1 represents the first multilayer perceptron, and Stoch_Pool represents the random pooling operation. The core representation o∈R is obtained by aggregating the representations of n sequences. d′ ; Merge the representations of the core and all sequences: F i =Repeat_Concat(S i-1 ,o i ) S i =MLP2(F i )+S i-1 The Repeat_Concat operation concatenates the core representation o into each sequence representation to obtain the feature F. i Then, a second multilayer perceptron (MLP2) is used to fuse the concatenated representations to obtain S. i The second multilayer perceptron MLP2:R d+d′ →R d Project the sequence representation back from the hidden dimension d+core dimension d′ of the concatenated sequence back to the hidden dimension d of the sequence; Step S3.4: Linear predictor, used to generate predicted power of photovoltaic power generation; Step S4: Add the predicted power back to the periodic component to obtain the final predicted power generation.

2. 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 periodic pattern is directly and explicitly modeled. Given multiple channels with prior period lengths, learnable cyclic periods are first generated. These cyclic periods are shared globally within the channels. By performing cyclic replication, cyclic components of the time series with the same length are obtained. The cyclic periods are trained together with the backbone module used for time series prediction through gradient backpropagation to generate a learned representation that reveals the cyclic pattern inside the sequence. The period 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 periodicity of the data is further examined using an autocorrelation function to measure the correlation between the time series and its lag value, as shown in the following formula: Where N represents the total number of observations, x t This represents the value of the time series at time t, where k represents the lag time. The autocorrelation function represents the mean 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 value. The largest peak value corresponds to the lag, which is consistent with the length of the longest period 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 performed based on the data analysis results, including replacing and filling out outliers and missing values ​​using linear interpolation; for time data, it is directly filled in according to the actual time. Consistency checks identify and remove data that is outside the reasonable range.

5. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: The periodic pattern in step S1 has the following characteristics: Regularity: Periodic patterns are characterized by data exhibiting similar fluctuation patterns repeatedly within certain time intervals; Stability: If the source of periodicity is stable, the periodic pattern does not change significantly over time; Multiple periodicity: Some data exhibit multiple overlapping periodicities.

6. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: In step S2, the cyclic component c is removed from the input original time series. t-L+1:t The 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 the cyclic component c. t-L+1:t The equivalent subsequences are obtained by the following steps: Step S2.1: Obtain Q by shifting the position to the left by the cycle period Q using t mod W. (t) t represents time; Step S2.2: Repeat Q (t) Operation [L / W] times, and connected in series These two equivalent subsequences can be represented as: Where W represents the period length, H represents the prediction time window length, and L represents the input time series length.

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 used. For the residual components obtained in step S2, the sequence is scaled to unit variance with zero mean as the center, and then reverse normalized on the predicted sequence.

8. The photovoltaic power generation time series prediction method based on periodic modeling and channel interaction according to claim 1, characterized in that: In step S3.4, after channel interaction, a single-layer linear predictor is used to generate the predicted power of photovoltaic power generation. Where Linear represents a linear mapping function, S N This represents the channel interaction output sequence of the Nth layer.

Images