A photovoltaic power prediction method and system based on LDCA-CEEMDAN-BiLSTM

Abstract

A photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM, characterized in that it comprises the following steps: step S1: performing CEEMDAN decomposition on the original photovoltaic power signal to obtain a plurality of IMF components; step S2: constructing an independent parallel prediction sub-model for each IMF component and its corresponding covariates, the covariates including one or more of irradiance, temperature, and time characteristics, and the structure of the parallel prediction sub-model being: Input→LDCA→BiLSTM→Dense; wherein the input is spliced from the current IMF component and the covariates; LDCA is a lightweight dynamic convolution attention module, used for extracting local time sequence features and generating dynamic attention weights; the present application realizes the cooperative optimization of signal denoising and feature enhancement, and significantly improves the prediction robustness of the model in the scene of severe irradiation fluctuation; by constructing an independent LDCA-BiLSTM sub-model for each IMF component, the fluctuation characteristics of different time scales can be modeled specifically, avoiding mutual interference of multi-scale features.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a photovoltaic power prediction method and system based on LDCA-CEEMDAN-BiLSTM. Background Technology

[0002] The penetration rate of photovoltaic power generation in new power systems is continuously increasing, but its output is affected by multiple factors such as irradiance, cloud cover, and temperature, exhibiting strong volatility, intermittency, and non-stationarity. High-precision short-term power forecasting is crucial for grid security, renewable energy integration, and market dispatch.

[0003] Current mainstream methods fall into two categories: physical model-based and data-driven. The latter, with its powerful nonlinear fitting capabilities, has become a research hotspot, with typical models including LSTM, GRU, and Transformer. To further improve performance, some methods introduce signal decomposition techniques such as CEEMD and VMD, modeling multi-scale components separately before fusing and predicting. However, existing technologies still face three major bottlenecks: 1. Lack of adaptability in denoising: Traditional methods (such as moving average and wavelet threshold) have fixed parameters, making it difficult to balance noise suppression and preservation of real fluctuations, especially when there are sudden changes in irradiance or rapid cloud movement. 2. Ignoring the discrete effects of weather types: Power response patterns differ significantly under different weather conditions (sunny, cloudy, rainy, etc.), but existing models often treat weather as a continuous feature without explicitly modeling its nonlinear and discrete characteristics, leading to inaccurate predictions when weather changes suddenly. 3. Decomposition and prediction are disconnected: In serial processes such as CEEMDAN+BiLSTM, signal decomposition is independent of the prediction target, sub-models are isolated from each other, the fusion strategy is simple, and it is impossible to dynamically weight or capture the nonlinear interaction between components.

[0004] While some studies have introduced attention mechanisms (such as SE Block and CBAM), they generally suffer from problems such as large parameter counts, reliance on global pooling, and high computational overhead, making them difficult to deploy on edge devices and failing to effectively integrate temporal local dependencies. Therefore, there is an urgent need for a lightweight, dynamically sensing, and end-to-end optimizable attention mechanism to coordinate signal decomposition and bidirectional temporal modeling, uniformly achieving adaptive denoising, weather category identification, and dynamic weighting of IMF components, thereby improving the accuracy of photovoltaic power generation forecasting. Summary of the Invention

[0005] To address the aforementioned shortcomings, the present invention aims to propose a photovoltaic power prediction method and system based on LDCA-CEEMDAN-BiLSTM, thereby solving the problem of inaccurate photovoltaic power generation prediction in existing methods.

[0006] To achieve this objective, the present invention adopts the following technical solution: a photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM, comprising the following steps: Step S1: Perform CEEMDAN decomposition on the original photovoltaic power signal to obtain several IMF components; Step S2: Construct an independent parallel prediction sub-model for each IMF component and its corresponding covariates. The covariates include one or more of irradiance, temperature, and time features. The structure of the parallel prediction sub-model is: Input→LDCA→BiLSTM→Dense. The input is composed of the current IMF component and covariates; LDCA is a lightweight dynamic convolutional attention module used to extract local temporal features and generate dynamic attention weights; BiLSTM has 128 neurons and is used to capture long-range temporal dependencies. Step S3: Weighted fusion of the single-component predicted values ​​of all outputs of the prediction sub-model is performed to obtain the total predicted value. The total predicted value is then denormalized to restore it to the original data scale, and the photovoltaic power prediction result is output.

[0007] Preferably, the LDCA processing in step S2 includes the following steps: Step A1: Input Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; Step A2: Process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; Step A3: Transpose attention map A back to its original time dimension to obtain the second processing value. ; Step A4: Process the second value Perform LayerNorm normalization to obtain the third processed value. ; Step A5: Third Processing Value Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

[0008] Preferably, the LDCA has no fully connected layers, but generates dynamic channel weights only through one-dimensional convolutional layers to achieve adaptive recalibration of input features.

[0009] Preferably, in step S2, LDCA is independently embedded in the parallel prediction sub-model corresponding to each IMF component to achieve component-level dynamic weighting and noise suppression for each IMF component.

[0010] Preferably, the LDCA module is placed at the front end of the BiLSTM network, and generates the attention weights by leveraging the complete context provided by the bidirectional temporal modeling capability of BiLSTM.

[0011] A photovoltaic power prediction system based on LDCA-CEEMDAN-BiLSTM, using the aforementioned photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM, includes: The signal decomposition module is used to perform CEEMDAN decomposition on the raw photovoltaic power signal and output multiple IMF components; The parallel sub-model module contains parallel prediction sub-models matching the number of IMF components. Each prediction sub-model integrates an LDCA module and a 128-neuron BiLSTM network. The input is the concatenated features of IMF components and covariates, and the output is the single-component prediction value. The fusion output module is used to perform weighted summation of all single-component prediction values ​​and inverse normalization processing to output the final photovoltaic power prediction result.

[0012] Preferably, the signal decomposition module includes a first submodule, a second submodule, a third submodule, a fourth submodule, and a fifth submodule; The first submodule is used to process the input. Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; The second submodule is used to process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; The third submodule is used to transpose attention map A back to the original time dimension to obtain the second processed value. ; The fourth submodule is used for processing the second value. Perform LayerNorm normalization to obtain the third processed value. ; The fifth submodule is used for the third processing value. Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

[0013] One of the above technical solutions has the following advantages or beneficial effects: the present invention achieves synergistic optimization of signal denoising and feature enhancement, which significantly improves the prediction robustness of the model in the scenario of severe irradiation fluctuations; by constructing an independent LDCA-BiLSTM sub-model for each IMF component, the fluctuation features at different time scales can be modeled in a targeted manner, avoiding mutual interference between multi-scale features. Attached Figure Description

[0014] Figure 1 This is a flowchart of one embodiment of the method of the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of one embodiment of the system of the present invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] like Figures 1-2 As shown, a photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM includes the following steps: Step S1: Perform CEEMDAN decomposition on the original photovoltaic power signal to obtain several IMF components (eigenmode function components). Step S2: Construct an independent parallel prediction sub-model for each IMF component and its corresponding covariates. The covariates include one or more of irradiance, temperature, and time features. The structure of the parallel prediction sub-model is: Input→LDCA→BiLSTM→Dense. The input is composed of the current IMF component and covariates; LDCA is a lightweight dynamic convolutional attention module used to extract local temporal features and generate dynamic attention weights; BiLSTM has 128 neurons and is used to capture long-range temporal dependencies. Step S3: Weighted fusion of all single-component predicted values ​​from the prediction sub-models is performed to obtain the total predicted value. This total predicted value is then denormalized to restore the original data scale, and the photovoltaic power prediction result is output. The formula for obtaining the photovoltaic power prediction result is as follows: Where M is the total number of IMF components. This represents the predicted value of the i-th single component of the prediction sub-model.

[0020] Firstly, addressing the lack of adaptability in traditional denoising methods, this invention employs CEEMDAN (Adaptive Complete Ensemble Empirical Mode Decomposition) to preprocess the original photovoltaic power signal. This decomposition technique can adaptively decompose non-stationary, highly fluctuating photovoltaic power sequences into several eigenmode function components (IMF components) based on the characteristics of the signal itself. Compared to methods with fixed parameters such as moving average or wavelet threshold, CEEMDAN can better handle nonlinear and non-stationary signals when irradiance changes abruptly or clouds move rapidly, automatically separating the fluctuation characteristics at different time scales in the original signal, thereby achieving adaptive noise suppression and preservation of true fluctuations.

[0021] To address the disconnect between decomposition and prediction in existing technologies, this innovative solution constructs an independent parallel prediction sub-model for each IMF component and embeds LDCA (Lightweight Dynamic Convolutional Attention Module) into the input of each sub-model. This allows the attention mechanism to directly apply to the decomposed signal components and their covariates. LDCA extracts local temporal features through lightweight convolution operations and dynamically generates attention weights for each IMF component. This design breaks the limitation of independent decomposition and prediction in the traditional CEEMDAN+BiLSTM serial process, achieving end-to-end joint optimization from signal decomposition to feature weighting and then to temporal modeling. LDCA can adaptively adjust weights according to the dynamic changes of each IMF component, capturing the potential nonlinear interactions between different components and avoiding information loss caused by simple fusion strategies. Addressing the difficulty in modeling the discrete impact of weather types, although the solution does not explicitly input weather type labels, it concatenates weather-related covariates such as irradiance and temperature with the IMF components and inputs them into the LDCA-BiLSTM network. The LDCA module can dynamically perceive the importance of these covariates in the current time series context, while BiLSTM captures the complex mapping relationship between irradiance changes and power output through bidirectional time series modeling. In fact, the model can automatically learn the nonlinear response law of irradiance, temperature and other features and photovoltaic power under different weather patterns during the training process. When encountering sudden weather changes, the dynamic weights generated by LDCA will be adjusted accordingly, enabling the model to have the ability to dynamically perceive weather changes.

[0022] This invention achieves synergistic optimization of signal denoising and feature enhancement, significantly improving the model's prediction robustness in scenarios with severe irradiation fluctuations. By constructing an independent LDCA-BiLSTM sub-model for each IMF component, the fluctuation features at different time scales can be modeled in a targeted manner, avoiding mutual interference between multi-scale features.

[0023] Preferably, the LDCA processing in step S2 includes the following steps: Step A1: Input Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; Step A2: Process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; Step A3: Transpose attention map A back to its original time dimension to obtain the second processing value. ; Step A4: Process the second value Perform LayerNorm normalization to obtain the third processed value. ; Step A5: Third Processing Value Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

[0024] First, step A1 transposes the input tensor from shape (B,T,C) to (B,C,T). This transpose operation places the time dimension last, which facilitates the subsequent one-dimensional convolution to extract local temporal features along the time axis, while maintaining the channel dimension as the channel input for the convolution, laying the foundation for subsequent cross-channel information interaction. Step A2 uses a one-dimensional convolutional layer with no bias, a single filter, and a kernel size of 3 to convolve the transposed first processed value. This convolutional layer slides along the time dimension, using a kernel of size 3 to capture the local dependencies of three adjacent time steps in each channel. Since the number of convolutional kernels is only 1 and there is no bias, the number of parameters is extremely low, achieving a truly lightweight design. At the same time, the single filter means that the generated attention map A has shape. That is, only one scalar attention value is generated at each time step. This attention value integrates the information of all channels within the local temporal window, thereby achieving adaptive extraction of temporal local features and avoiding the information loss and high computational cost caused by global pooling in traditional attention mechanisms. Step A3 transposes the attention map A back to the original time dimension to obtain the shape (B,T,1), so that it can be aligned with the original input (B,T,C) in the time dimension, which is convenient for subsequent element-wise multiplication. Step A4 performs LayerNorm normalization on the transposed attention map. This normalization operation is performed along the time... Standardizing the attention values ​​in the time step dimension can stabilize the distribution of attention weights, prevent activation values ​​from being too large or too small, thereby accelerating training convergence and enhancing the model's generalization ability. Step A5 activates the normalized third processing value through the Sigmoid function, mapping the attention weights to the (0,1) interval to obtain the final dynamic attention weights. Then, the weights are multiplied element-wise with the original input to achieve adaptive weighted enhancement of the input features along the time dimension, that is, giving greater weight to important time steps and suppressing minor or noisy time steps, thereby completing the dynamic enhancement output of features.

[0025] Preferably, the LDCA has no fully connected layers and generates dynamic channel weights only through one-dimensional convolutional layers to achieve adaptive recalibration of input features. The LDCA is a lightweight module without fully connected layers or biases, configured with only a single-filter one-dimensional convolutional layer, making it suitable for deployment on resource-constrained edge devices.

[0026] Preferably, in step S2, LDCA is independently embedded in the parallel prediction sub-model corresponding to each IMF component to achieve component-level dynamic weighting and noise suppression for each IMF component.

[0027] Because CEEMDAN decomposes the original non-stationary photovoltaic power sequence into several IMF components from high frequency to low frequency, where the high-frequency components mainly contain random noise and rapid fluctuations, while the low-frequency components reflect trend and regular changes, if a uniform attention mechanism is used to process all components in the same way, it is impossible to optimize for the unique time series patterns of each component. However, independently embedding LDCA allows each sub-model to learn an attention weighting strategy specific to its own input component based on the statistical characteristics and dynamic changes of that component. Specifically, for the high-frequency IMF component containing a large amount of noise, its corresponding LDCA module dynamically generates... The attention weights can adaptively identify and suppress the moments corresponding to random noise, while preserving the true fluctuation characteristics during sudden changes in irradiance. For low-frequency IMF components that reflect stable trends, LDCA focuses on strengthening periodic patterns and long-range dependencies. The advantage of this component-level dynamic weighting is that the parameters of each LDCA module are optimized only for a specific IMF component, avoiding cross-component feature interference. This allows the model to capture power response patterns at different time scales more accurately. At the same time, the lightweight design of each LDCA module ensures that even if multiple IMF components are deployed in parallel, the overall computational burden will not increase significantly.

[0028] Preferably, the LDCA module is placed at the front end of the BiLSTM network, and generates the attention weights by leveraging the complete context provided by the bidirectional temporal modeling capability of BiLSTM.

[0029] A photovoltaic power prediction system based on LDCA-CEEMDAN-BiLSTM, using the aforementioned photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM, includes: The signal decomposition module is used to perform CEEMDAN decomposition on the raw photovoltaic power signal and output multiple IMF components; The parallel sub-model module contains parallel prediction sub-models matching the number of IMF components. Each prediction sub-model integrates an LDCA module and a 128-neuron BiLSTM network. The input is the concatenated features of IMF components and covariates, and the output is the single-component prediction value. The fusion output module is used to perform weighted summation of all single-component prediction values ​​and inverse normalization processing to output the final photovoltaic power prediction result.

[0030] Preferably, the signal decomposition module includes a first submodule, a second submodule, a third submodule, a fourth submodule, and a fifth submodule; The first submodule is used to process the input. Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; The second submodule is used to process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; The third submodule is used to transpose attention map A back to the original time dimension to obtain the second processed value. ; The fourth submodule is used for processing the second value. Perform LayerNorm normalization to obtain the third processed value. ; The fifth submodule is used for the third processing value. Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

[0031] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0032] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM, characterized in that, Includes the following steps: Step S1: Perform CEEMDAN decomposition on the original photovoltaic power signal to obtain several IMF components; Step S2: Construct an independent parallel prediction sub-model for each IMF component and its corresponding covariates. The covariates include one or more of irradiance, temperature, and time features. The structure of the parallel prediction sub-model is: Input→LDCA→BiLSTM→Dense. The input is composed of the current IMF component and covariates; LDCA is a lightweight dynamic convolutional attention module used to extract local temporal features and generate dynamic attention weights; BiLSTM has 128 neurons and is used to capture long-range temporal dependencies. Step S3: Weighted fusion of the single-component predicted values ​​of all outputs of the prediction sub-model is performed to obtain the total predicted value. The total predicted value is then denormalized to restore it to the original data scale, and the photovoltaic power prediction result is output.

2. The photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM according to claim 1, characterized in that, The LDCA processing procedure in step S2 includes the following steps: Step A1: Input Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; Step A2: Process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; Step A3: Transpose attention map A back to its original time dimension to obtain the second processing value. ; Step A4: Process the second value Perform LayerNorm normalization to obtain the third processed value. ; Step A5: Third Processing Value Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

3. A photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM according to claim 1 or 2, characterized in that, The LDCA has no fully connected layers; it generates dynamic channel weights only through one-dimensional convolutional layers to achieve adaptive recalibration of input features.

4. The photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM according to claim 1, characterized in that, In step S2, LDCA is independently embedded in the parallel prediction sub-model corresponding to each IMF component to achieve component-level dynamic weighting and noise suppression for each IMF component.

5. The photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM according to claim 2, characterized in that, The LDCA module is placed at the front end of the BiLSTM network and generates the attention weights by leveraging the complete context provided by the bidirectional temporal modeling capability of BiLSTM.

6. A photovoltaic power prediction system based on LDCA-CEEMDAN-BiLSTM, using the photovoltaic power prediction method based on LDCA-CEEMDAN-BiLSTM as described in any one of claims 1 to 5, characterized in that, include: The signal decomposition module is used to perform CEEMDAN decomposition on the raw photovoltaic power signal and output multiple IMF components; The parallel sub-model module contains parallel prediction sub-models matching the number of IMF components. Each prediction sub-model integrates an LDCA module and a 128-neuron BiLSTM network. The input is the concatenated features of IMF components and covariates, and the output is the single-component prediction value. The fusion output module is used to perform weighted summation of all single-component prediction values ​​and inverse normalization processing to output the final photovoltaic power prediction result.

7. A photovoltaic power prediction system based on LDCA-CEEMDAN-BiLSTM according to claim 6, wherein the signal decomposition module comprises a first submodule, a second submodule, a third submodule, a fourth submodule, and a fifth submodule; The first submodule is used to process the input. Transpose the values ​​along the channel and time dimensions to obtain the first processed value. Where B is the batch size, T is the time step, and C is the number of feature channels; The second submodule is used to process the first value using an unbiased, single-filter one-dimensional convolutional layer. Perform convolution to obtain attention map A, wherein the kernel size of the one-dimensional convolutional layer is 3; The third submodule is used to transpose attention map A back to the original time dimension to obtain the second processed value. ; The fourth submodule is used for processing the second value. Perform LayerNorm normalization to obtain the third processed value. ; The fifth submodule is used for the third processing value. Attention weights are obtained by activation using the Sigmoid function. The attention weights With input Element-wise multiplication outputs enhanced features. .

Images