Ultra-short-term photovoltaic power prediction method and device fusing ICEEMDAN and improved GRU-ECA

Abstract

The invention relates to an ultra-short-term photovoltaic power prediction method fusing ICEEMDAN and improved GRU-ECA. The method comprises the steps of collecting photovoltaic power and associated data and performing preprocessing; obtaining a plurality of groups of IMF components; the GRU-ECA prediction model is improved, the multiple groups of IMF components are input into the improved GRU-ECA prediction model, and a photovoltaic power prediction sequence is obtained; and evaluating the prediction sequence of the photovoltaic power by adopting a multi-dimensional error index. According to the method, the original photovoltaic power sequence is decomposed into multiple groups of IMF components with different frequencies through ICEEMDAN, the calculation cost is reduced while key features are reserved, and the process efficiency is improved; the improved ECA module strengthens the importance of key features such as solar irradiance and environment temperature through one-dimensional global average pooling and adaptive channel weight learning; the method can effectively adapt to the dynamic change of photovoltaic power input characteristics, simplifies the model structure, improves the calculation efficiency, and can accurately capture key information, thereby achieving the efficient and high-precision prediction of the ultra-short-term photovoltaic power.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to an ultra-short-term photovoltaic power prediction method and device that integrates ICEEMDAN and an improved GRU-ECA. Background Technology

[0002] Photovoltaic power generation is highly constrained by natural factors. Instantaneous changes in sunlight intensity, fluctuations in ambient temperature, and cloud movement all contribute to the strong intermittency, volatility, and randomness of photovoltaic power output. This characteristic not only results in low photovoltaic power absorption rates and frequent curtailment, but also poses significant challenges to grid load balancing and the coordinated control of energy storage devices, seriously affecting the safe and stable operation of the power grid. Accurate ultra-short-term photovoltaic power forecasting is a key technical means to solve these problems. Currently, mainstream photovoltaic power forecasting methods are mainly divided into three categories:

[0003] Physical method: Based on the geographical location of the photovoltaic power station, photovoltaic module parameters (such as photoelectric conversion efficiency and temperature coefficient), and meteorological forecast data, the power is calculated through a photoelectric conversion physical model. However, this method requires extremely high accuracy of meteorological data, the modeling process is complex, and the prediction error increases significantly under extreme weather conditions, limiting its practicality.

[0004] Statistical methods, including time series analysis and multiple linear regression, predict power by mining statistical patterns in historical data. These methods are computationally efficient and have a fast response time, but they struggle to handle the nonlinear characteristics of photovoltaic power and exhibit poor robustness in the face of sudden changes in sunlight.

[0005] Machine learning and deep learning methods: Traditional machine learning methods (such as support vector machines) can handle some nonlinear data, but they are sensitive to outliers; existing deep learning methods (such as LSTM and CNN-LSTM) have improved the extraction of time-series features, but they cannot fully adapt to the strong fluctuations of photovoltaic power, and they do not pay enough attention to key features (such as solar irradiance), so the prediction accuracy is difficult to meet engineering requirements. Summary of the Invention

[0006] To address the problems of complex photovoltaic power prediction decomposition processes, redundant model structures, and difficulty in balancing accuracy and computational efficiency in existing technologies, the primary objective of this invention is to provide an ultra-short-term photovoltaic power prediction method that integrates ICEEMDAN and the improved GRU-ECA, while retaining key features, reducing computational costs, improving process efficiency, and achieving efficient and high-precision prediction of ultra-short-term photovoltaic power.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an ultra-short-term photovoltaic power prediction method integrating ICEEMDAN and an improved GRU-ECA, the method comprising the following sequential steps:

[0008] (1) Collect photovoltaic power and related data, and preprocess the collected data to obtain the preprocessed photovoltaic power sequence;

[0009] (2) The preprocessed photovoltaic power sequence was decomposed using ICEEMDAN to obtain multiple IMF components;

[0010] (3) Improve the GRU-ECA prediction model to obtain the improved GRU-ECA prediction model. Input multiple sets of IMF components into the improved GRU-ECA prediction model to obtain the prediction sequence of photovoltaic power.

[0011] (4) The prediction sequence of photovoltaic power is evaluated using multi-dimensional error indices.

[0012] Step (1) specifically includes the following steps:

[0013] (1a) Collect photovoltaic power and related data: Collect photovoltaic power data of photovoltaic power station at a fixed cycle of once every 15 minutes within a time range of 1 to 3 years, and collect solar irradiance and ambient temperature data at the corresponding time.

[0014] (1b) Preprocessing the collected data: Linear interpolation is used to fill missing values, that is, the missing data is replaced by the average value of the data in the previous and next collection cycles at the missing time; then, Min-Max normalization is performed to obtain the preprocessed photovoltaic power sequence.

[0015] Step (2) specifically includes the following steps:

[0016] (2a) The preprocessed photovoltaic power sequence is decomposed using ICEEMDAN, specifically including the following steps:

[0017] (2a1) Let For the preprocessed photovoltaic power sequence, special noise is removed. Add to middle:

[0018] ;

[0019] In the formula: To add the first Photovoltaic power sequence after special noise; The expected signal-to-noise ratio for the first decomposition iteration; For Empirical Mode Decomposition First-order modal components; Indicates the first step used for auxiliary decomposition Group white noise;

[0020] (2a2) Calculate the residuals of the first decomposition. and use minus First-order modal components are obtained :

[0021] ;

[0022] ;

[0023] In the formula: Operators for generating local mean values ​​of signals; The number of white noise groups;

[0024] (2a3) Continue adding white noise signal and calculate the residual of the first decomposition. and second-order modal components :

[0025] ;

[0026] ;

[0027] In the formula, The expected signal-to-noise ratio for the second decomposition iteration; These are the second-order modal components generated by empirical mode decomposition;

[0028] (2a4) Calculate the residual of the k-th decomposition. and k-th order modal components :

[0029] ;

[0030] ;

[0031] In the formula, The residual of the (k-1)th decomposition, Let be the expected signal-to-noise ratio for the k-th decomposition iteration;

[0032] (2a5) Repeat step (2a4) until the maximum number of iterations is reached or the residual can no longer be decomposed, then proceed to step (2b).

[0033] (2b) Retain all the IMF components obtained from the decomposition to obtain multiple sets of IMF components.

[0034] Step (3) specifically includes the following steps:

[0035] (3a) Improve the ECA module to obtain the improved ECA module:

[0036] The two-dimensional global average pooling layer in the ECA module is replaced with a one-dimensional global average pooling layer, and an interactive coverage k is adaptively determined based on the size of the input channel C.

[0037] ;

[0038] In the formula: t is an intermediate calculated value, Let be the odd number closest to t; C is the current total number of channels; γ and b are fixed values ​​of 2 and 1 respectively. After adaptively determining the value of k, all channels share the learned attention weight parameters, then the i-th channel... weight Represented as:

[0039] ;

[0040] In the formula: To and Adjacent A collection of channels; Weight parameters shared by all channels; It is the sigmoid function; This is the global average feature value corresponding to the i-th channel;

[0041] Finally, multiple sets of IMF components were combined with By combining these, we obtain a weighted sequence of feature maps. Important channel features are assigned larger weights to enhance them, while ineffective channel features are assigned smaller weights to suppress them.

[0042] (3b) Constructing the improved GRU-ECA prediction model: The weighted sequence output by the improved ECA module is used as the basis for the prediction model. Input the GRU module to capture the long-term temporal dependencies of each IMF component; output the photovoltaic power prediction value for the next 0 to 4 hours through the fully connected layer, and restore the actual magnitude through Min-Max inverse normalization;

[0043] (3c) Combine all IMF components with solar irradiance and ambient temperature data as input to the improved GRU-ECA prediction model to obtain the prediction sequence of photovoltaic power.

[0044] Step (4) specifically refers to: calculating the mean absolute error index. Root mean square error index and mean absolute percentage error The value is used to evaluate the error of the photovoltaic power prediction sequence. , and The smaller the value, the higher the prediction accuracy.

[0045] Another object of the present invention is to provide an electronic device comprising:

[0046] Processor; and

[0047] A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the ultra-short-term photovoltaic power prediction method as described above, which integrates ICEEMDAN and the improved GRU-ECA.

[0048] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the ultra-short-term photovoltaic power prediction method as described above, which integrates ICEEMDAN and the improved GRU-ECA.

[0049] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, the present invention simplifies the data processing flow by using ICEEMDAN single decomposition. ICEEMDAN decomposes the original photovoltaic power sequence into multiple groups of IMF components with different frequencies, eliminating the need for secondary decomposition and clustering operations. This reduces computational costs and improves process efficiency while retaining key features. Second, the improved ECA module in the present invention strengthens the importance of key features such as solar irradiance and ambient temperature through one-dimensional global average pooling and adaptive channel weight learning. The GRU module efficiently captures the long-term temporal dependencies of photovoltaic power, such as intraday periodicity and seasonal trends, through a gating structure. The synergistic effect of the two modules can effectively adapt to the dynamic changes in photovoltaic power input features. While simplifying the model structure and improving computational efficiency, it can still accurately capture key information, thereby achieving efficient and high-precision prediction of ultra-short-term photovoltaic power. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0051] like Figure 1 As shown, an ultra-short-term photovoltaic power prediction method and device integrating ICEEMDAN and an improved GRU-ECA are disclosed. The method includes the following sequential steps:

[0052] (1) Collect photovoltaic power and related data, and preprocess the collected data to obtain the preprocessed photovoltaic power sequence;

[0053] (2) The preprocessed photovoltaic power sequence was decomposed using ICEEMDAN, i.e., improved complete set empirical mode decomposition, to obtain multiple sets of IMF components;

[0054] (3) Improve the GRU-ECA prediction model to obtain the improved GRU-ECA prediction model. Input multiple sets of IMF components into the improved GRU-ECA prediction model to obtain the prediction sequence of photovoltaic power.

[0055] (4) The prediction sequence of photovoltaic power is evaluated using multi-dimensional error indices.

[0056] Step (1) specifically includes the following steps:

[0057] (1a) Collect photovoltaic power and related data: Collect photovoltaic power data of photovoltaic power station at a fixed cycle of once every 15 minutes within a time range of 1 to 3 years, and collect solar irradiance and ambient temperature data at the corresponding time.

[0058] (1b) Preprocessing the collected data: Linear interpolation is used to fill missing values, that is, the missing data is replaced by the average value of the data in the previous and next collection cycles at the missing time; then, Min-Max normalization is performed to obtain the preprocessed photovoltaic power sequence.

[0059] Step (2) specifically includes the following steps:

[0060] (2a) The preprocessed photovoltaic power sequence is decomposed using ICEEMDAN, specifically including the following steps:

[0061] (2a1) Let For the preprocessed photovoltaic power sequence, special noise is removed. Add to middle:

[0062] ;

[0063] In the formula: To add the first Photovoltaic power sequence after special noise; The expected signal-to-noise ratio for the first decomposition iteration; For Empirical Mode Decomposition First-order modal components; Indicates the first step used for auxiliary decomposition Group white noise;

[0064] (2a2) Calculate the residuals of the first decomposition. and use minus First-order modal components are obtained :

[0065] ;

[0066] ;

[0067] In the formula: Operators for generating local mean values ​​of signals; The number of white noise groups;

[0068] (2a3) Continue adding white noise signal and calculate the residual of the first decomposition. and second-order modal components :

[0069] ;

[0070] ;

[0071] In the formula, The expected signal-to-noise ratio for the second decomposition iteration; These are the second-order modal components generated by empirical mode decomposition;

[0072] (2a4) Calculate the residual of the k-th decomposition. and k-th order modal components :

[0073] ;

[0074] ;

[0075] In the formula, The residual of the (k-1)th decomposition, Let be the expected signal-to-noise ratio for the k-th decomposition iteration;

[0076] (2a5) Repeat step (2a4) until the maximum number of iterations is reached or the residual can no longer be decomposed, then proceed to step (2b).

[0077] (2b) Retain all the IMF components obtained from the decomposition to obtain multiple sets of IMF components.

[0078] Step (3) specifically includes the following steps:

[0079] (3a) Improve the ECA module to obtain the improved ECA module:

[0080] The two-dimensional global average pooling layer in the ECA module is replaced with a one-dimensional global average pooling layer to compress time-series data of different dimensions. The improved ECA module can operate on a large number of features extracted by the TCN module (temporal convolutional network) and assign weights to different features. Finally, the original input features are combined with the channel weights to obtain a weighted sequence of feature maps. Important channel features are assigned larger weights to enhance them, while ineffective channel features are assigned smaller weights to suppress them.

[0081] The two-dimensional global average pooling layer in the ECA module is replaced with a one-dimensional global average pooling layer, and an interactive coverage k is adaptively determined based on the size of the input channel C.

[0082] ;

[0083] In the formula: t is an intermediate calculated value, Let be the odd number closest to t; C is the current total number of channels; γ and b are fixed values ​​of 2 and 1 respectively. After adaptively determining the value of k, all channels share the learned attention weight parameters, then the i-th channel... weight Represented as:

[0084] ;

[0085] In the formula: To and Adjacent A collection of channels; Weight parameters shared by all channels; It is the sigmoid function; This is the global average feature value corresponding to the i-th channel;

[0086] Finally, multiple sets of IMF components were combined with By combining these, we obtain a weighted sequence of feature maps. Important channel features are assigned larger weights to enhance them, while ineffective channel features are assigned smaller weights to suppress them.

[0087] (3b) Constructing the improved GRU-ECA prediction model: The weighted sequence output by the improved ECA module is used as the basis for the prediction model. Input the GRU module to capture the long-term temporal dependencies of each IMF component; output the photovoltaic power prediction value for the next 0 to 4 hours through the fully connected layer, and restore the actual magnitude through Min-Max inverse normalization;

[0088] (3c) Combine all IMF components with solar irradiance and ambient temperature data as input to the improved GRU-ECA prediction model to obtain the photovoltaic power prediction sequence and realize high-precision prediction of ultra-short-term photovoltaic power.

[0089] Step (4) specifically refers to: calculating the mean absolute error index. Root mean square error index and mean absolute percentage error The value is used to evaluate the error of the photovoltaic power prediction sequence. , and The smaller the value, the higher the prediction accuracy.

[0090] In summary, this invention simplifies the data processing flow by employing ICEEMDAN single-pass decomposition. ICEEMDAN decomposes the original photovoltaic power sequence into multiple IMF components of different frequencies, eliminating the need for secondary decomposition and clustering operations. This reduces computational costs and improves process efficiency while preserving key features. The improved ECA module in this invention strengthens the importance of key features such as solar irradiance and ambient temperature through one-dimensional global average pooling and adaptive channel weight learning. The GRU module efficiently captures long-term temporal dependencies of photovoltaic power, such as intraday periodicity and seasonal trends, through a gating structure. The synergistic effect of these two modules effectively adapts to dynamic changes in photovoltaic power input characteristics. While simplifying the model structure and improving computational efficiency, it still accurately captures key information, thereby achieving efficient and high-precision prediction of ultra-short-term photovoltaic power.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for predicting ultra-short-term photovoltaic power by integrating ICEEMDAN and an improved GRU-ECA, characterized in that: The method includes the following steps in sequence: (1) Collect photovoltaic power and related data, and preprocess the collected data to obtain the preprocessed photovoltaic power sequence; (2) The preprocessed photovoltaic power sequence was decomposed using ICEEMDAN to obtain multiple IMF components; (3) Improve the GRU-ECA prediction model to obtain the improved GRU-ECA prediction model. Input multiple sets of IMF components into the improved GRU-ECA prediction model to obtain the prediction sequence of photovoltaic power. (4) The prediction sequence of photovoltaic power is evaluated using multi-dimensional error indices.

2. The ultra-short-term photovoltaic power prediction method integrating ICEEMDAN and the improved GRU-ECA as described in claim 1, characterized in that: Step (1) specifically includes the following steps: (1a) Collect photovoltaic power and related data: Collect photovoltaic power data of photovoltaic power station at a fixed cycle of once every 15 minutes within a time range of 1 to 3 years, and collect solar irradiance and ambient temperature data at the corresponding time. (1b) Preprocessing the collected data: Linear interpolation is used to fill missing values, that is, the missing data is replaced by the average value of the data in the previous and next collection cycles at the missing time; then, Min-Max normalization is performed to obtain the preprocessed photovoltaic power sequence.

3. The ultra-short-term photovoltaic power prediction method integrating ICEEMDAN and the improved GRU-ECA as described in claim 1, characterized in that: Step (2) specifically includes the following steps: (2a) The preprocessed photovoltaic power sequence is decomposed using ICEEMDAN, specifically including the following steps: (2a1) Let For the preprocessed photovoltaic power sequence, special noise is removed. Add to middle: ； In the formula: To add the first Photovoltaic power sequence after special noise; The expected signal-to-noise ratio for the first decomposition iteration; For Empirical Mode Decomposition First-order modal components; Indicates the first step used for auxiliary decomposition Group white noise; (2a2) Calculate the residuals of the first decomposition. and use minus First-order modal components are obtained : ； ； In the formula: Operators for generating local mean values ​​of signals; The number of white noise groups; (2a3) Continue adding white noise signal and calculate the residual of the first decomposition. and second-order modal components : ； ； In the formula, The expected signal-to-noise ratio for the second decomposition iteration; These are the second-order modal components generated by empirical mode decomposition; (2a4) Calculate the residual of the k-th decomposition. and k-th order modal components : ； ； In the formula, The residual of the (k-1)th decomposition, Let be the expected signal-to-noise ratio for the k-th decomposition iteration; (2a5) Repeat step (2a4) until the maximum number of iterations is reached or the residual can no longer be decomposed, then proceed to step (2b). (2b) Retain all the IMF components obtained from the decomposition to obtain multiple sets of IMF components.

4. The ultra-short-term photovoltaic power prediction method integrating ICEEMDAN and the improved GRU-ECA as described in claim 1, characterized in that: Step (3) specifically includes the following steps: (3a) Improve the ECA module to obtain the improved ECA module: The two-dimensional global average pooling layer in the ECA module is replaced with a one-dimensional global average pooling layer, and an interactive coverage k is adaptively determined based on the size of the input channel C. ； In the formula: t is an intermediate calculated value, Let be the odd number closest to t; C is the current total number of channels; γ and b are fixed values ​​of 2 and 1 respectively. After adaptively determining the value of k, all channels share the learned attention weight parameters, then the i-th channel... weight Represented as: ； In the formula: To and Adjacent A collection of channels; Weight parameters shared by all channels; It is the sigmoid function; This is the global average feature value corresponding to the i-th channel; Finally, multiple sets of IMF components were combined with By combining these, we obtain a weighted sequence of feature maps. Important channel features are assigned larger weights to enhance them, while ineffective channel features are assigned smaller weights to suppress them. (3b) Constructing the improved GRU-ECA prediction model: The weighted sequence output by the improved ECA module is used as the basis for the prediction model. Input the GRU module to capture the long-term temporal dependencies of each IMF component; output the photovoltaic power prediction value for the next 0 to 4 hours through the fully connected layer, and restore the actual magnitude through Min-Max inverse normalization; (3c) Combine all IMF components with solar irradiance and ambient temperature data as input to the improved GRU-ECA prediction model to obtain the prediction sequence of photovoltaic power.

5. The ultra-short-term photovoltaic power prediction method integrating ICEEMDAN and the improved GRU-ECA according to claim 1, characterized in that: Step (4) specifically refers to: calculating the mean absolute error index. Root mean square error index and mean absolute percentage error The value is used to evaluate the error of the photovoltaic power prediction sequence. , and The smaller the value, the higher the prediction accuracy.

6. An electronic device, comprising: processor; as well as A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the ultra-short-term photovoltaic power prediction method that integrates ICEEMDAN and the improved GRU-ECA as described in any one of claims 1-5.

7. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the ultra-short-term photovoltaic power prediction method as described in any one of claims 1-5.

Images