Photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network

By improving the combination of empirical mode decomposition and optimizing long short-term memory networks, the problem of poor adaptability of nonlinear and non-stationary characteristics in photovoltaic power prediction is solved, achieving higher accuracy and stable medium-term prediction, adapting to photovoltaic data in different climate zones, and supporting power system dispatching and photovoltaic power plant operation and maintenance.

CN121863356APending Publication Date: 2026-04-14HUZHOU ELECTRIC POWER SUPPLY CO OF STATE GRID ZHEJIANG ELECTRIC POWER CO LTD +3
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511928730.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-19
Publication Date
2026-04-14

AI Technical Summary

Technical Problem

In existing photovoltaic power prediction technologies, traditional models are poorly adapted to the nonlinear and non-stationary characteristics of photovoltaic power sequences and cannot effectively capture power fluctuations caused by meteorological factors such as radiance and temperature. Conventional empirical mode decomposition (EMD) suffers from endpoint effects and mode aliasing, which affect prediction accuracy. The key parameters of long short-term memory networks (LSTM) rely on manual tuning, which is inefficient and prone to getting trapped in local optima, resulting in weak model generalization ability.

Method used

By employing an improved Empirical Mode Decomposition (EE-ANEMD) combined with adaptive noise processing and endpoint extension, the Long Short-Term Memory (LSTM) network is optimized. The number of hidden layer units and learning rate are optimized through an improved Sparrow Search Algorithm (ISSA). This constructs a multi-scale feature adaptive decomposition and temporal prediction model, improving the accuracy and stability of mid-term predictions.

Benefits of technology

It significantly improves the accuracy and stability of medium-term photovoltaic power forecasts, reduces the root mean square error of the test set, enhances the ability to capture power fluctuation trends caused by meteorological factors, and improves the model's generalization ability and training efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121863356A_ABST
    Figure CN121863356A_ABST
Patent Text Reader

Abstract

The invention discloses a photovoltaic power prediction method based on improved empirical mode decomposition and an optimized long short-term memory network, and the method comprises the steps: firstly carrying out the preprocessing of abnormal value elimination, missing value filling, normalization and the like of photovoltaic power and related meteorological data, and improving the data quality; then, an improved empirical mode decomposition (EE-ANEMD) algorithm is adopted to decompose the preprocessed power sequence into a multi-scale intrinsic mode function component and a residual term, and high-frequency noise, intermediate-frequency fluctuation and a low-frequency trend are effectively separated; global optimization is carried out on the hidden layer unit number, the initial learning rate and the maximum number of training times of the LSTM network through an improved sparrow search algorithm (ISSA), finally, the optimized LSTM is utilized to carry out training prediction on each component, and results are fused and subjected to reverse normalization to obtain a final value. Experiments show that the test set RMSE of the method is reduced compared with that of a single LSTM, the mid-term prediction precision is remarkably improved, and reliable technical support is provided for power system dispatching, new energy consumption planning and photovoltaic power station operation and maintenance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network. Background Technology

[0002] In the global transition towards a cleaner and lower-carbon energy structure, the development and utilization of renewable energy has become a core measure to address climate change and the energy crisis. Among these, solar photovoltaic (PV) power generation, with its abundant resources and zero carbon emissions, has become a crucial pillar of the global energy transition. In recent years, global PV installation capacity has experienced explosive growth, posing unique challenges to the stable operation of the power system due to the large-scale grid connection of PV power generation. Because PV output power is highly susceptible to natural factors such as solar irradiance, ambient humidity, and cloud cover, its output exhibits strong volatility and randomness. This characteristic can lead to abnormal fluctuations in grid voltage and frequency. With the continuous increase in PV penetration, extreme cases may trigger grid imbalances, even causing large-scale blackouts, seriously threatening the safe and reliable operation of the power system.

[0003] Against this backdrop, accurate forecasting of photovoltaic (PV) power has become a key technology for ensuring grid stability and optimizing power dispatch. Based on the forecast period, PV power forecasting can be categorized into four types: ultra-short-term, short-term, medium-term, and long-term. Among these, medium-term forecasting is of significant guiding importance for power sectors in formulating weekly power generation plans, coordinating conventional and renewable energy output, and developing equipment operation and maintenance plans. However, current industry research largely focuses on ultra-short-term and short-term forecasting, with relatively few technical solutions for medium-term forecasting. Furthermore, existing methods struggle to meet the accuracy and stability requirements of practical applications. Current mainstream PV power forecasting methods are mainly divided into three categories: physical models, statistical models, and artificial intelligence models, but all have significant limitations. Physical models predict photovoltaic (PV) power output by constructing the physical equations of PV module photoelectric conversion. This requires precise knowledge of module parameters and meteorological data, but their high complexity and computational cost make them unsuitable for diverse real-world scenarios. Statistical models, such as ARIMA and SARIMA, predict power output by mining statistical patterns from historical data, but they are poorly adapted to nonlinear and non-stationary PV power sequences, resulting in significant prediction errors. Artificial intelligence (AI) models, such as neural networks and support vector machines, are widely used in prediction due to their powerful nonlinear fitting capabilities. Long Short-Term Memory (LSTM) networks, in particular, excel in PV prediction because they effectively handle long-range dependencies in time-series data. However, single AI models are susceptible to noise in the raw data and struggle to fully extract multi-scale features, limiting accuracy improvements in medium-term predictions.

[0004] To address these issues, combined models have become a research hotspot, aiming to improve prediction performance by integrating the advantages of different models. Empirical Mode Decomposition (EMD), as an adaptive signal decomposition method, can decompose complex nonlinear and non-stationary sequences into a series of intrinsic mode functions (IMFs), effectively separating different frequency components in the data and providing higher-quality input features for subsequent modeling. Combining EMD with LSTM is expected to fully leverage the former's advantages in feature decomposition and the latter's advantages in time series prediction, improving the accuracy and stability of medium-term photovoltaic power prediction. However, the original EMD suffers from endpoint effects and mode aliasing. Endpoint effects cause the decomposed IMFs to be distorted at the beginning and end of the sequence, increasing the data deviation at the beginning and end of subsequent predictions. Mode aliasing mixes high-frequency noise of photovoltaic power (such as instantaneous cloud cover) with mid-frequency fluctuations (such as daytime power gradient changes) in the same IMF, causing LSTM to be unable to accurately extract features at each scale, further reducing the accuracy of medium-term predictions. Meanwhile, the combination of EMD and LSTM still suffers from problems such as insufficient model adaptability and poor parameter optimization, and a mature technical solution has not yet been developed for medium-term prediction of optical power. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing photovoltaic (PV) power prediction technologies. Traditional models have poor adaptability to the nonlinear and non-stationary characteristics of PV power sequences, failing to effectively capture power fluctuations caused by meteorological factors such as radiance and temperature. Conventional Empirical Mode Decomposition (EMD) suffers from significant endpoint effects and mode aliasing, leading to component distortion and affecting subsequent prediction accuracy. Key parameters of Long Short-Term Memory (LSTM) networks (such as the number of hidden layer units and learning rate) rely on manual tuning, which is not only inefficient but also prone to getting trapped in local optima, resulting in weak model generalization ability. This invention proposes a PV power prediction method based on improved Empirical Mode Decomposition and optimized LSTM networks. It integrates Improved Empirical Mode Decomposition (EE-ANEMD), Improved Sparrow Search Algorithm (ISSA), and LSTM networks. Through adaptive and accurate decomposition of multi-scale features, global parameter optimization, and collaborative design of time-series prediction models, it comprehensively improves the accuracy and stability of medium-term PV power prediction, ultimately providing reliable technical support for power system dispatch planning, new energy consumption planning optimization, and PV power plant operation and maintenance strategy adjustment.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: a photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network, comprising: Historical photovoltaic power data and related meteorological data of photovoltaic power plants are collected and processed to obtain photovoltaic power time-series data and related meteorological characteristic data. Based on endpoint extension and adaptive noise improvement, empirical mode decomposition of photovoltaic power time series data yields K IMF components and a residual term. With the goal of minimizing the global optimal root mean square error, the number of hidden layer units, the maximum number of training rounds, and the initial learning rate of the Long Short-Term Memory Network are optimized through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation, resulting in an optimized Long Short-Term Memory Network. A dataset is constructed by integrating correlated meteorological feature data and IMF components. An optimized long short-term memory network is trained on the dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted by the trained optimized long short-term memory network.

[0007] Preferably, the step of collecting and processing historical photovoltaic power data and associated meteorological data from the photovoltaic power station to obtain photovoltaic power time-series data includes: The collected historical photovoltaic power data and related meteorological data of the photovoltaic power station include photovoltaic power and corresponding radiance, temperature, air pressure and humidity arranged in chronological order; Use 3 The criteria identify and remove outlier data that deviate from the mean by three times the standard deviation; missing values ​​are supplemented by linear interpolation of power data at adjacent time points; the data after removing outliers and supplementing missing values ​​are mapped to the [0,1] interval for standardization to obtain photovoltaic power time series data and associated meteorological feature data.

[0008] The above standardization process improves data quality, providing a standardized and highly reliable input foundation for subsequent feature decomposition and model training.

[0009] Preferably, the empirical mode decomposition of photovoltaic power time series data based on endpoint extension and adaptive noise improvement yields K IMF components and a residual term, including: The photovoltaic power sequence data is symmetrically copied into two 10% length samples. The first 10% of the samples are flipped to form the front extension segment, and the last 10% of the samples are flipped to form the back extension segment, thus forming a complete photovoltaic power sequence. To address the issue of incomplete information and few extreme points at the beginning and end of photovoltaic power sequences, which leads to distorted upper and lower envelope fitting during EMD decomposition and affects decomposition accuracy, the aforementioned "mirror extension method" is used to extend the sequence before decomposition. After decomposition, the extended portion is truncated, which allows the envelope fitting at the boundary to be based on more complete neighborhood information, reducing endpoint errors from the source.

[0010] Extended photovoltaic power sequence Add Gaussian white noise with an amplitude of 0.1% of the original data's standard deviation. Constructing a noisy photovoltaic power sequence As shown below: The empirical mode decomposition of the noisy photovoltaic power sequence is shown below: After decomposition, the extended part is truncated, and only the decomposition result corresponding to the original sequence is retained, resulting in K eigenmode functions and one residual term.

[0011] To address the issue that photovoltaic power is affected by multiple factors such as solar radiation and temperature, high-frequency fluctuations (such as power jumps caused by instantaneous cloud cover) and mid-frequency trends (such as changes in daytime power gradients) are prone to superimpose, resulting in mode aliasing. This causes the decomposed IMF components to contain multiple frequency components, making it impossible to accurately reflect a single physical process. Based on the EMD algorithm, the aforementioned "adaptive noise-assisted decomposition" method is introduced. The uniform frequency of the noise can dilute the energy coupling between modes, allowing high-frequency components (such as instantaneous disturbances) and mid-frequency components (such as the influence of weather systems) to be naturally separated, ensuring that each IMF component corresponds to only a single-scale fluctuation characteristic.

[0012] Preferably, the termination condition for empirical mode decomposition of the noisy photovoltaic power sequence is set as follows: Set the IMF component obtained from the k-th screening. The calculated energy The sum of the squares of all data points for this component and the number of extreme points for this component, and the energy of the original noisy photovoltaic power sequence. This is the sum of squares of the noisy photovoltaic power data; , where N is the sequence length.

[0013] The screening process will automatically terminate when two consecutive screening results simultaneously meet the following two conditions: The ratio of the energy difference of the IMF component obtained from two consecutive screenings to the energy of the original noisy photovoltaic power sequence is less than a set threshold, as shown below: Number of extreme points of IMF components obtained from two consecutive screenings The difference is less than a set threshold, as shown below: .

[0014] The original EMD algorithm relies on fixed empirical thresholds to determine whether IMF components meet decomposition conditions (e.g., default screening terminates after 20 times) when processing photovoltaic power data. This approach is poorly adapted to complex fluctuations in photovoltaic power, potentially leading to incomplete component separation due to insufficient screening or the introduction of redundant components due to over-screening, thus affecting the input quality of subsequent prediction models. To address this issue, the screening termination condition is improved by adopting the aforementioned "energy difference dual-threshold termination condition." This condition adaptively matches the intensity of photovoltaic power fluctuations, ensuring that the decomposed IMF components are free of redundancy.

[0015] Preferably, the optimized Long Short-Term Memory (LSTM) network, which aims to minimize the global optimal root mean square error, optimizes the number of hidden layer units, the maximum number of training epochs, and the initial learning rate through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation, to obtain the optimized LSM network, includes: Set the range of values ​​for the number of hidden layer units, the range of values ​​for the maximum number of training epochs, and the range of values ​​for the initial learning rate; An initial population with a normal distribution and truncated boundaries is generated based on the center values ​​of the ranges for the number of hidden layer units, the maximum number of training epochs, and the initial learning rate, ensuring that all initial solutions fall within a preset reasonable range, as shown below: Where X(i,j) represents the candidate position of the i-th individual in the ISSA on the j-th dimension parameter, that is, the value of the j-th parameter in the i-th group of LSTM hyperparameter candidate combinations (not the final optimal solution, but only the initial candidate value). randn() It is a random number that follows a standard normal distribution. ub(j), lb(j) It represents the search range for parameter j.

[0016] Calculate the fitness value of the initial population and proceed with the iteration. In each iteration, the adaptive weight is linearly decayed based on the fitness value of the previous generation population to balance global exploration and local exploitation, as shown below: in, Indicates the first t The adaptive weights in the next iteration are used to balance global exploration and local development. This represents the initial weights, used in the early stages of iteration to enhance global exploration capabilities and cover the complex parameter space of photovoltaic data. The final weights are used in later iterations to enhance local development capabilities and optimize LSTM hyperparameters. t represents the current iteration number, with a value range of... , Indicates the maximum number of iterations; Based on linearly decaying adaptive weights, the population is updated according to roles: the discoverer dynamically selects exponential or linear exploration according to the safety threshold, the first half of the joiner adaptively follows the global optimal solution and the second half is randomly reborn, and the vigilant updates its position by using Logistic chaotic perturbation or random normal perturbation according to the relationship between the individual fitness and the current global optimal solution. Recalculate the updated fitness value of the population when the number of iterations reaches [number missing]. When the maximum number of iterations is reached, iteration stops, and the individual corresponding to the minimum global root mean square error among the fitness values ​​of the population after all generations of updates is the optimal individual. The number of hidden layer units, the maximum number of training epochs, and the initial learning rate corresponding to the current optimal individual are used as the final hyperparameters of LSTM.

[0017] Preferably, the population update is based on linearly decaying adaptive weights and assigned to different roles: discoverers dynamically select exponential or linear exploration methods according to a safety threshold; joiners adaptively follow the global optimum in the first half and randomly respawn in the second half; and vigilant individuals update their positions using either Logistic chaotic perturbation or random normal perturbation based on the relationship between their individual fitness and the current global optimum. The discoverer is set to dynamically select the exploration method based on the safety threshold ST. When the safety status is: The discovery exponential exploration method is used to expand the scope. The update formula for the discovery exponential exploration is as follows: in ,express t+1 The position of the i-th discoverer in the next iteration; This represents the current position of the i-th discoverer at the t-th iteration; i represents the discoverer index (value range: ...). (pop is the population size, PD is the proportion of discoverers). The first number is a random number that follows a uniform distribution in the range [0,1]; ST is the safety threshold, used to determine whether the environment is safe. The current situation is safe; the exploration area needs to be expanded. A random number that follows a uniform distribution in [0,1], used for switching between safety and alert settings; This represents the adaptive weight in the t-th iteration. The maximum number of iterations is represented by an exponential function. To achieve distance-dependent exploration attenuation, discoverers with smaller index i have smaller exploration amplitudes, while discoverers with larger index i have larger exploration amplitudes, adapting to the multi-peak and nonlinear hyperparameter space characteristics of photovoltaic data.

[0018] when (In alert state), a linear exploration fine-grained search is employed, with the following formula: Define the number of new members (1-PD): For the latter half of the individuals (i>pop / 2), a random regeneration formula is used as follows. Its core is to avoid population stagnation and solve the problem of rigid follower behavior in the basic SSA algorithm. For the first half of the individuals, an adaptive following method is adopted (adjusting the direction based on the optimal solution). The formula is as follows. Its core is to explore near the high-quality area and balance diversity and search efficiency. in, ⊙ represents the element-wise product, dim refers to the dimension of the hyperparameter to be optimized. The optimization parameters of LSTM are the number of hidden layer units, the initial learning rate, and the maximum number of training iterations, a total of 3 parameters, so dim=3. U is a random vector, dim=3, so U is a 1x3 random vector; A is a random matrix with dimension 1×dim, ub is the upper bound of the hyperparameter values, and lb is the lower bound of the hyperparameter values. Define a vigilant (SD=0.2): Introduce chaotic perturbations to escape local optima when an individual's fitness > the current optimum (getting stuck in a local state), as shown in the following formula: Where the Logistic chaotic mapping function The iterative formula for generating chaotic sequences to escape local optima is as follows: k=1,2,3 j=1,2,...,dim in, Indicates the first t+1 During the nth iteration, the 1st idx The updated location of each vigilant; 0.1 represents the global optimal position in the t-th iteration (the current optimal combination of LSTM hyperparameters); 0.1 represents the random perturbation coefficient (controlling the strength of the normal distribution perturbation to avoid excessive perturbation that deviates from the optimal region); represents a random vector that follows a standard normal distribution N(0,1) (dimension dim=3, introducing randomness to maintain population diversity); 0.05 represents the chaos perturbation coefficient (controlling the degree of influence of the chaotic sequence, balancing "escaping local" and "searching for stability"); This indicates that the signal was generated based on the current location of the vigilant. Logistic Chaotic sequence (dimension dim=3); This represents the j-th element in the k-th iteration of the chaotic sequence (k=3 is the iteration number, empirically set to ensure chaotic characteristics). C 0 ( j ) represents the initial value of the chaotic sequence; idx Indicates the index of the vigilant (randomly selected from the population, number of which is...). (SD=0.2 represents the proportion of vigilant individuals).

[0019] When individual fitness The current optimal (stable search) formula is as follows: random perturbation is the current position plus 0.2 multiplied by the normal perturbation. The core is to maintain population diversity and avoid premature convergence during stable search. Through global optimal guidance, normal random perturbation, Logistic Combinatorial strategies of chaotic perturbations, with Based on this, ensure that the update direction is biased towards high-quality areas; normal perturbation Maintaining population diversity to prevent premature convergence of the algorithm; chaotic sequences By leveraging its ergodicity and aperiodicity, it helps the vigilant escape local optima.

[0020] In photovoltaic (PV) forecasting research, PV power output is dynamically affected by multiple factors such as radiance, temperature, air pressure, and humidity. Its time series exhibits strong nonlinearity, non-stationarity, and multi-scale fluctuations, making it extremely sensitive to the hyperparameters of the forecasting model (LSTM). This is specifically reflected in the following aspects: Too few hidden units in an LSTM will lead to underfitting and an inability to capture complex features, while too many will lead to overfitting and sensitivity to noise. Too few maximum training epochs will lead to insufficient model training, while too many will increase computational costs and may cause overfitting. An excessively large initial learning rate can lead to training oscillations and non-convergence, while an excessively small initial learning rate results in extremely slow convergence.

[0021] Therefore, photovoltaic (PV) forecasting requires an efficient and robust optimization algorithm capable of searching for globally optimal hyperparameters to adapt to the complexity of the data. The basic Sparrow Search Algorithm (SSA), using uniform random initialization, is prone to initial solution convergence, easily misses globally optimal regions, lacks an adaptive mechanism, has fixed weights during the search process, suffers from insufficient early global exploration, weak local exploration capabilities in the later stages, and suffers from simple update of the watchdog and insufficient search. This invention proposes the Improved Sparrow Search Algorithm (ISSA), which addresses the nonlinear and non-stationary characteristics of PV data, solving the problems of insufficient diversity, susceptibility to local optima, and insufficient search inherent in the basic SSA and traditional algorithms. It provides a more accurate and robust hyperparameter optimization scheme for LSTM models. The basic SSA algorithm uses a uniform random initialization method, which can lead to initial solutions being concentrated in local regions. This is especially problematic when the photovoltaic data parameter space is complex, as it can easily miss the global optimum. The ISSA algorithm proposes a normal distribution + boundary truncation initialization method, ensuring that the initial solutions are normally distributed around the center of the parameter space. Simultaneously, max / min truncation ensures that the solutions remain within the legal range, significantly improving population diversity and covering more potential optimal regions, making it particularly suitable for the multi-peak parameter space of photovoltaic data. In the basic SSA (Sparrow Search Algorithm), the search process weights are fixed, leading to insufficient early global exploration and excessive later local exploitation. This fails to adapt to the multi-scale characteristics of photovoltaic data, ranging from macroscopic trends to microscopic fluctuations. To address this, a linear decay adaptive weight and role-based update strategy is proposed, where weights decay linearly with the number of iterations, balancing exploration and exploitation.

[0022] Preferably, the dataset is constructed based on associated meteorological feature data and IMF components. The optimized long short-term memory network is trained using this dataset. The trained optimized long short-term memory network then predicts each IMF component and residual term. The predicted values ​​of each component are superimposed and inversely normalized to obtain the final photovoltaic power prediction result, including: Each IMF component is fused with associated meteorological feature data, and input-output sample pairs are constructed according to a preset delay step size to ensure the temporal correlation between the input features and the prediction target. The input features contain two core pieces of information: one part is the historical time series value of the corresponding IMF component, which is used to capture the dynamic change pattern of the component itself; the other part is the associated meteorological feature data synchronized with the timestamp of the historical time series value, which is used to introduce the impact of external meteorological factors on photovoltaic power. The reconstructed sample pairs are randomly shuffled to avoid model training bias caused by the continuity of sample time. They are then divided into training and test sets according to a preset ratio. The training set is used for model parameter learning, and the test set is used to verify the model's generalization performance. Based on the optimized Long Short-Term Memory (LSTM) network, an independent optimized LSTM prediction model is constructed for each IMF component and residual term. The mean squared error is used as the loss function for model training. Using the optimized long short-term memory network models that have been trained, predictions are made for the corresponding IMF components and residual terms, and prediction results for each component and residual term are obtained. The core advantage of using component prediction is that it transforms the original nonlinear and non-stationary photovoltaic power sequence into multiple simple and stationary subsequences, which are modeled separately. This greatly reduces the learning difficulty of a single model and improves the prediction accuracy of each subsequence.

[0023] The predicted values ​​of all IMF components and residual terms are superimposed, and then restored to the original dimensions through inverse normalization to obtain the final photovoltaic power prediction result.

[0024] This invention also provides a photovoltaic power prediction system based on improved empirical mode decomposition and optimized long short-term memory network, comprising: The data acquisition module is used to collect and process historical photovoltaic power data and related meteorological data from photovoltaic power plants to obtain photovoltaic power time-series data and related meteorological characteristic data. An improved empirical mode decomposition module is used to perform empirical mode decomposition on photovoltaic power time series data based on endpoint extension and adaptive noise improvement, to obtain K IMF components and a residual term; The Long Short-Term Memory (LSTM) network parameter optimization module is used to optimize the number of hidden layer units, maximum training epochs, and initial learning rate of the LSM network with the goal of minimizing the global optimal root mean square error. This is achieved through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation. A dataset is constructed by integrating correlated meteorological feature data and IMF components. An optimized long short-term memory network is trained on the dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted by the trained optimized long short-term memory network.

[0025] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described above.

[0026] The present invention also provides an electronic device, including a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described above are implemented.

[0027] The beneficial effects of this invention are as follows: The improved Empirical Mode Decomposition (EE-ANEMD) achieves accurate separation of multi-scale features. It combines endpoint extension to suppress endpoint effects and adaptive noise injection to mitigate mode aliasing. Furthermore, it incorporates an optimized Long Short-Term Memory (LSTM) network using an improved Sparrow Search Algorithm (ISSA), significantly improving prediction accuracy compared to a single LSTM model. The root mean square error of the test set is reduced, and the model demonstrates stronger ability to capture power fluctuation trends caused by meteorological factors such as radiance and temperature in medium-term predictions, effectively enhancing the accuracy and stability of photovoltaic power prediction. Simultaneously, EE-ANEMD inherits the advantage of traditional EMD, which does not require preset basis functions. It can adaptively decompose the nonlinear and non-stationary characteristics of photovoltaic power and automatically selects effective components through energy difference thresholds, further improving decomposition quality. ISSA, through population diversification initialization, adaptive weight adjustment, and chaotic perturbation strategies, achieves global optimization of LSTM parameters, completely eliminating reliance on human experience. This allows the model to maintain stable performance on photovoltaic data from different climate zones in the Northern Hemisphere, significantly enhancing its generalization ability.

[0028] Furthermore, by employing a component prediction approach, complex sequences are decomposed into stationary components and modeled separately, significantly reducing the learning difficulty of a single model and improving training efficiency. Compared with deep learning models of equivalent accuracy, this method can provide more reliable prediction results for weekly power system scheduling, helping to reduce photovoltaic curtailment rates. At the same time, it provides accurate data support for energy storage system configuration, reducing energy storage capacity requirements. This has significant practical implications for the efficient operation of photovoltaic power plants, the consumption of new energy sources, and the safe and stable operation of the power system, and possesses extremely high engineering application value.

[0029] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0030] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 Here is a flowchart of a photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network provided in Example 1; Figure 2 Here is a flowchart of the improved empirical mode decomposition photovoltaic power timing data in Example 1; Figure 3 This is a diagram showing the decomposition results of photovoltaic power time series data improved by empirical mode decomposition in Example 1; Figure 4 This is a flowchart of the improved sparrow search algorithm for optimizing the long short-term memory network in Example 1; Figure 5 This is a block diagram of a photovoltaic power prediction system based on improved empirical mode decomposition and optimized long short-term memory network, as provided in Example 2. Detailed Implementation

[0031] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0032] Example 1 like Figure 1 As shown, this embodiment provides a photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network, including the following steps: S1. Collect historical photovoltaic power data and related meteorological data of the photovoltaic power station, including radiance, temperature, air pressure, and humidity. Process the collected data to obtain photovoltaic power time-series data and related meteorological characteristic data, as follows: Outlier removal: through 3 The criteria identify and remove outlier data that deviates from the mean by three times the standard deviation. Missing value imputation: Missing values ​​are filled by linear interpolation of power data from adjacent time points; Normalization: The data after removing outliers and filling in missing values ​​are mapped to the [0,1] interval for standardization to obtain photovoltaic power time series data and associated meteorological feature data.

[0033] S2, such as Figure 2 As shown, based on the improved empirical mode decomposition (EE-ANEMD) photovoltaic power time series data, K IMF components and one residual term are obtained, as follows: 2.1 The photovoltaic power sequence data is symmetrically copied into two 10% length samples. The first 10% of the samples are flipped to form the front extension segment, and the last 10% of the samples are flipped to form the back extension segment, thus forming a complete photovoltaic power sequence. 2.2 Extended Photovoltaic Power Sequence Add the amplitude as the standard deviation of the original data 0.1% Gaussian white noise Constructing a noisy photovoltaic power sequence As shown below: In the formula, N represents a normal distribution; 2.3 Empirical mode decomposition is performed on the noisy photovoltaic power sequence, as shown below: The conditions for terminating the decomposition are set as follows: Set the IMF component obtained from the k-th screening. The calculated energy The sum of the squares of all data points for this component and the number of extreme points for this component, and the energy of the original noisy photovoltaic power sequence. This is the sum of squares of the noisy photovoltaic power data; , where N is the sequence length.

[0034] The screening process will automatically terminate when two consecutive screening results simultaneously meet the following two conditions: The ratio of the energy difference of the IMF component obtained from two consecutive screenings to the energy of the original noisy photovoltaic power sequence is less than a set threshold, as shown below: Number of extreme points of IMF components obtained from two consecutive screenings The difference is less than a set threshold, as shown below: .

[0035] After decomposition, the extended portion is truncated, retaining only the decomposition result corresponding to the original sequence, such as... Figure 3 As shown, it includes the original power sequence, five intrinsic mode functions (IMF1-IMF5, corresponding to high-frequency noise, second-high frequency disturbance, mid-frequency fluctuation, second-low frequency trend, and low-frequency fundamental fluctuation, respectively) and one residual term (reflecting the overall power change trend).

[0036] S3. With the goal of minimizing the global optimal root mean square error, the number of hidden layer units, maximum training epochs, and initial learning rate of the Long Short-Term Memory (LSTM) network are optimized using a sparrow search algorithm improved by population initialization, adaptive weights, and a watchdog perturbation. The optimized LSM network is obtained as follows: Figure 4 As shown: 3.1 Set the range of values ​​for the number of hidden layer units, the maximum number of training epochs, and the initial learning rate; 3.2 Based on the center values ​​of the ranges for the number of hidden layer units, the maximum number of training epochs, and the initial learning rate, an initial population with a normal distribution and truncated boundaries is generated, as shown below: Where X(i,j) represents the candidate position of the i-th individual in the ISSA on the j-th dimension parameter, that is, the value of the j-th parameter in the i-th group of LSTM hyperparameter candidate combinations (not the final optimal solution, but only the initial candidate value). randn() It is a random number that follows a standard normal distribution. ub(j), lb(j) It represents the search range for parameter j.

[0037] 3.2 Calculate the fitness value of the initial population and proceed with the iteration. In each iteration, the adaptive weight is linearly decayed based on the fitness value of the previous generation population, as shown below: in, Indicates the first t The adaptive weights in the next iteration are used to balance global exploration and local development. This represents the initial weights, used in the early stages of iteration to enhance global exploration capabilities and cover the complex parameter space of photovoltaic data. The final weights are used in later iterations to enhance local development capabilities and optimize LSTM hyperparameters. t represents the current iteration number, with a value range of... , Indicates the maximum number of iterations; The basic SSA algorithm has a small search range, especially for the high-dimensional parameter space of photovoltaic data, where it is almost impossible to find the global optimum. In the ISSA algorithm designed in this embodiment, a larger population size and number of iterations are used, and the fitness of all individuals is forcibly recalculated after each iteration, and the fitness changes of each generation are printed.

[0038] 3.3 Based on linearly decaying adaptive weights, population updates are performed according to roles, as shown below: The discoverer is set to dynamically select the exploration method based on the safety threshold ST. When the safety status is: The discovery exponential exploration method is used to expand the scope. The update formula for the discovery exponential exploration is as follows: in ,express t+1 The position of the i-th discoverer in the next iteration; This represents the current position of the i-th discoverer at the t-th iteration; i represents the discoverer index (value range: ...). (pop is the population size, PD is the proportion of discoverers). The first number is a random number that follows a uniform distribution in the range [0,1]; ST is the safety threshold, used to determine whether the environment is safe. The current situation is safe; the exploration area needs to be expanded. A random number that follows a uniform distribution in [0,1], used for switching between safety and alert settings; This represents the adaptive weight in the t-th iteration. The maximum number of iterations is represented by an exponential function. To achieve distance-dependent exploration attenuation, discoverers with smaller index i have smaller exploration amplitudes, while discoverers with larger index i have larger exploration amplitudes, adapting to the multi-peak and nonlinear hyperparameter space characteristics of photovoltaic data.

[0039] when (In alert state), a linear exploration fine-grained search is employed, with the following formula: Define the number of new members (1-PD): For the latter half of the individuals (i>pop / 2), a random regeneration formula is used as follows. Its core is to avoid population stagnation and solve the problem of rigid follower behavior in the basic SSA algorithm. For the first half of the individuals, an adaptive following method is adopted (adjusting the direction based on the optimal solution). The formula is as follows. Its core is to explore near the high-quality area and balance diversity and search efficiency. in, ⊙ represents the element-wise product, dim refers to the dimension of the hyperparameter to be optimized. The optimization parameters of LSTM are the number of hidden layer units, the initial learning rate, and the maximum number of training iterations, a total of 3 parameters, so dim=3. U is a random vector, dim=3, so U is a 1x3 random vector; A is a random matrix with dimension 1×dim, ub is the upper bound of the hyperparameter values, and lb is the lower bound of the hyperparameter values. Define a vigilant (SD=0.2): Introduce chaotic perturbations to escape local optima when an individual's fitness > the current optimum (getting stuck in a local state), as shown in the following formula: Where the Logistic chaotic mapping function The iterative formula for generating chaotic sequences to escape local optima is as follows: k=1,2,3 j=1,2,...,dim in, Indicates the first t+1 During the nth iteration, the 1st idx The updated location of each vigilant; 0.1 represents the global optimal position in the t-th iteration (the current optimal combination of LSTM hyperparameters); 0.1 represents the random perturbation coefficient (controlling the strength of the normal distribution perturbation to avoid excessive perturbation that deviates from the optimal region); represents a random vector that follows a standard normal distribution N(0,1) (dimension dim=3, introducing randomness to maintain population diversity); 0.05 represents the chaos perturbation coefficient (controlling the degree of influence of the chaotic sequence, balancing escape from local and search stability). This indicates that the signal was generated based on the current location of the vigilant. Logistic Chaotic sequence (dimension dim=3); This represents the j-th element in the k-th iteration of the chaotic sequence (k=3 is the iteration number, empirically set to ensure chaotic characteristics). C 0 ( j ) represents the initial value of the chaotic sequence; idx Indicates the index of the vigilant (randomly selected from the population, number of which is...). (SD=0.2 represents the proportion of vigilant individuals).

[0040] When individual fitness The current optimal (stable search) formula is as follows: random perturbation is the current position plus 0.2 multiplied by the normal perturbation. The core is to maintain population diversity and avoid premature convergence during stable search. Through global optimal guidance, normal random perturbation, Logistic Combinatorial strategies of chaotic perturbations, with Based on this, ensure that the update direction is biased towards high-quality areas; normal perturbation Maintaining population diversity to prevent premature convergence of the algorithm; chaotic sequences By leveraging its ergodicity and aperiodicity, it helps the vigilant escape local optima.

[0041] 3.4 Recalculate the fitness value of the updated population. When the number of iterations reaches the maximum number of iterations, output the population individual corresponding to the minimum global root mean square error among the fitness values ​​of the population after all generations of updates as the optimal population individual. The number of hidden layer units, the maximum number of training rounds, and the initial learning rate corresponding to the current optimal population individual are used as the final hyperparameters of LSTM.

[0042] S4. A dataset is constructed based on associated meteorological feature data and IMF components. An optimized Long Short-Term Memory (LSTM) network is trained using this dataset. The trained LSM network then predicts each IMF component and residual term. The predicted values ​​of each component are superimposed and inversely normalized to obtain the final photovoltaic power prediction result, specifically including: 4.1 Each IMF component is fused with associated meteorological feature data to construct input-output sample pairs, ensuring the temporal correlation between features and prediction targets. A time delay step is set, including a delay step of k=5, which uses relevant data from the previous 5 hours to predict the corresponding IMF component value for the 6th hour. This step size can cover the key influence period of short-term power fluctuations, meeting the requirements for extracting temporal features in medium-term forecasts. For each IMF component... The input features contain two core types of information: one is historical power features, specifically the power data from the previous 5 hours. Sequence value, specifically , , , , One type is used to capture the temporal variation pattern of this component; the other type is meteorological characteristics, which are related to... Real-time synchronized solar irradiance G(t), temperature T(t), air pressure P(t), and relative humidity H(t) are used to introduce the external environment that affects photovoltaic power.

[0043] 4.2 The reconstructed sample pairs are randomly shuffled to avoid model training bias caused by the continuity of sample time sequence. They are then divided into training set and test set according to a preset ratio. The training set is used for model parameter learning, and the test set is used to verify the generalization performance of the model. 4.3 Based on the optimized hyperparameters of the improved sparrow search algorithm, independent optimized Long Short-Term Memory (LSTM) network prediction models are constructed for each IMF component and residual term. The mean squared error is used as the loss function during model training. The LSTM prediction network consists of an input layer, an LSTM layer, a fully connected layer, a loss function, and an optimizer. The input layer primarily receives input features of different dimensions, with the feature dimensions consistent with the reconstructed dataset. The LSTM layer has a hidden layer unit count of [missing information]. (Optimal value obtained from ISSA optimization), the network return sequence mode is set to output only the result of the last time step, adapting to single-step prediction requirements; the fully connected layer has one neuron to output the corresponding IMF component. The system maps LSTM features to the predicted target. Specifically, the input layer receives feature tensors and feeds them into the network in batches. The LSTM layer operates with the ISSA-optimized number of hidden units, capturing temporal dependencies of components through a gating mechanism, and outputting only the high-dimensional features of the last time step. The fully connected layer, through a single neuron, linearly maps this high-dimensional feature to 1D, ultimately outputting the predicted value for the corresponding component. The mean squared error (MSE) between the predicted value and the true value in the training set is calculated. To minimize the MSE, backpropagation using the chain rule is employed to calculate the gradient of the loss function with respect to the trainable parameters (weights, biases) of the LSTM and fully connected layers, mitigating the vanishing gradient problem. Then, combined with the ISSA-optimized initial learning rate, the parameters are updated in the gradient descent direction, gradually reducing the MSE. This cycle of forward propagation outputting predicted values ​​→ calculating MSE → backpropagation calculating gradients → parameter updates is repeated until the termination condition is met: reaching the maximum number of training iterations optimized by ISSA, triggering an early stopping strategy, or the MSE drops to the convergence threshold. After training, the optimal parameter set corresponding to the minimum MSE is retained.

[0044] 4.4 Using the trained optimized long short-term memory network models, predictions are made based on the corresponding IMF components and residual terms to obtain the prediction results for each component and residual term; The predicted values ​​of all IMF components and residual terms are superimposed, and then restored to the original dimensions through inverse normalization to obtain the final photovoltaic power prediction result.

[0045] Example 2 like Figure 5 As shown, this embodiment provides a photovoltaic power prediction system based on improved empirical mode decomposition and optimized long short-term memory network, including: The data acquisition module is used to collect and process historical photovoltaic power data and related meteorological data from photovoltaic power plants to obtain photovoltaic power time-series data and related meteorological characteristic data. An improved empirical mode decomposition module is used to perform empirical mode decomposition on photovoltaic power time series data based on endpoint extension and adaptive noise improvement, to obtain K IMF components and a residual term; The Long Short-Term Memory (LSTM) network parameter optimization module is used to optimize the number of hidden layer units, maximum training epochs, and initial learning rate of the LSM network with the goal of minimizing the global optimal root mean square error. This is achieved through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation, resulting in an optimized LSM network. A dataset is constructed by integrating correlated meteorological feature data and IMF components. An optimized long short-term memory network is trained on the dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted by the trained optimized long short-term memory network.

[0046] Example 3 The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described in Example 1.

[0047] Example 4 The present invention also provides an electronic device, including a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described in Embodiment 1 are implemented.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network, characterized in that, include: Historical photovoltaic power data and related meteorological data of photovoltaic power plants are collected and preprocessed to obtain photovoltaic power time series data and related meteorological characteristic data. Based on endpoint extension and adaptive noise improvement, empirical mode decomposition of photovoltaic power time series data yields K IMF components and a residual term. With the goal of minimizing the global optimal root mean square error, the number of hidden layer units, the maximum number of training rounds, and the initial learning rate of the Long Short-Term Memory Network are optimized through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation, resulting in an optimized Long Short-Term Memory Network. A dataset is constructed by integrating correlated meteorological feature data and IMF components. An optimized long short-term memory network is trained on the dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted by the trained optimized long short-term memory network.

2. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 1, characterized in that, The process involves collecting and processing historical photovoltaic power data and associated meteorological data from photovoltaic power plants to obtain photovoltaic power time-series data, including: The collected historical photovoltaic power data and related meteorological data of the photovoltaic power station include photovoltaic power and corresponding radiance, temperature, air pressure and humidity arranged in chronological order; Adopt 3 The criteria identify and remove outlier data that deviates from the mean by more than a set amount; missing values ​​are supplemented by linear interpolation of power data at adjacent time points; the data after removing outliers and supplementing missing values ​​are mapped to the [0,1] interval for standardization to obtain photovoltaic power time series data and associated meteorological feature data.

3. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 1, characterized in that, The improved empirical mode decomposition (IMF) of photovoltaic power time series data based on endpoint extension and adaptive noise improvement yields K IMF components and a residual term, including: The photovoltaic power sequence data is symmetrically copied with a set length of samples. The copied sample at the beginning is flipped to become the front extension segment, and the copied sample at the end is flipped to become the back extension segment, thus forming a complete photovoltaic power sequence. Add Gaussian white noise with an amplitude within the standard deviation range of the original data to the extended photovoltaic power sequence to construct a noisy photovoltaic power sequence; Empirical mode decomposition is performed on the noisy photovoltaic power sequence. After the decomposition is completed, the extended part is truncated, and only the decomposition result corresponding to the original sequence is retained, resulting in K eigenmode functions and one residual term.

4. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 3, characterized in that, The termination condition for empirical mode decomposition of the noisy photovoltaic power sequence is set as follows: Define the IMF component obtained from the k-th screening, calculate the energy as the sum of the squares of all data points of this component and the number of extreme points of this component, and the energy of the original noisy photovoltaic power sequence is the sum of the squares of the noisy photovoltaic power data; The screening process will automatically terminate when two consecutive screening results simultaneously meet the following two conditions: The ratio of the energy difference of the IMF component obtained from two consecutive screenings to the energy of the original noisy photovoltaic power sequence is less than a set threshold. The difference in the number of extreme points of the IMF component obtained from two consecutive screenings is less than the set threshold.

5. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 1, characterized in that, The optimization of the Long Short-Term Memory (LSTM) network, which aims to minimize the global optimal root mean square error, utilizes a sparrow search algorithm improved through population initialization, adaptive weights, and a watchdog perturbation to optimize the number of hidden layer units, the maximum number of training epochs, and the initial learning rate. The resulting optimized LSM network includes: Set the range of values ​​for the number of hidden layer units, the range of values ​​for the maximum number of training epochs, and the range of values ​​for the initial learning rate; Based on the center values ​​of the ranges for the number of hidden layer units, the maximum number of training epochs, and the initial learning rate, an initial population with a normal distribution and truncated boundaries is generated, as shown below: Where X(i,j) represents the candidate position of the i-th individual in the ISSA on the j-th dimension parameter, that is, the value of the j-th parameter in the i-th group of LSTM hyperparameter candidate combinations; randn() It is a random number that follows a standard normal distribution. ub(j), lb(j) It is the search range for parameter j; Calculate the fitness value of the initial population and proceed with the iteration. In each iteration, the adaptive weight is linearly decreased based on the fitness value of the previous generation, as shown below: in, Indicates the first t Adaptive weights for the next iteration Indicates the initial weights. The final weight is represented by t, which represents the current iteration number and has a range of values. , Indicates the maximum number of iterations; Based on linearly decaying adaptive weights, the population is updated according to roles: the discoverer dynamically selects exponential or linear exploration according to the safety threshold, the first half of the joiner adaptively follows the global optimal solution and the second half is randomly reborn, and the vigilant updates its position by using Logistic chaotic perturbation or random normal perturbation according to the relationship between the individual fitness and the current global optimal solution. The fitness values ​​of the updated population are recalculated. When the number of iterations reaches the maximum number of iterations, the population individual corresponding to the minimum global root mean square error among all the updated population fitness values ​​is the optimal population individual. The number of hidden layer units, the maximum number of training rounds, and the initial learning rate corresponding to the current optimal population individual are used as the final hyperparameters of LSTM.

6. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 5, characterized in that, The adaptive weights based on linear decay are used for population updates based on roles: discoverers dynamically select exponential or linear exploration methods according to a safety threshold; newcomers adaptively follow the global optimum in the first half and randomly respawn in the second half; and vigilant individuals update their positions using either Logistic chaotic perturbation or random normal perturbation based on the relationship between their individual fitness and the current global optimum. The discoverer is set to dynamically select the exploration method based on the safety threshold ST. When the safety status is: The discovery exponential exploration method is used to expand the scope. The update formula for the discovery exponential exploration is as follows: in ,express t+1 The position of the i-th discoverer in the next iteration; This represents the current position of the i-th discoverer at the t-th iteration; i represents the discoverer index. is a random number that follows a uniform distribution in [0,1]; ST is the safety threshold; These are random numbers that follow a uniform distribution in the range [0,1]. This represents the adaptive weight in the t-th iteration. This indicates the maximum number of iterations. when A linear exploration fine search is used, with the following formula: The joining party is configured to randomly respawn the latter half of the individuals, using the following formula. For the first half of the individuals, adaptive following is used, as shown in the following formula: in, ⊙ represents the element-wise product, dim refers to the dimension of the hyperparameter to be optimized, U is a random vector, if dim=3, U is a 1x3 random vector; A is a random matrix of dimension 1×dim, ub is the upper bound of the hyperparameter value, and lb is the lower bound of the hyperparameter value. A watchdog is introduced to create chaotic perturbations that allow individuals to escape local optima when their fitness is greater than the current optimum. The formula is as follows: Where the Logistic chaotic mapping function The iterative formula is as follows: ,k=1,2,3 ,j=1,2,...,dim in, Indicates the first t+1 During the nth iteration, the 1st idx The updated location of each vigilant; This represents the globally optimal position in the t-th iteration; Let N(0,1) represent a random vector that follows a standard normal distribution. This indicates that the signal was generated based on the current location of the vigilant. Logistic Chaotic sequence; This represents the j-th element in the k-th iteration of the chaotic sequence; C 0 ( j ) represents the initial value of the chaotic sequence; idx Index indicating vigilant individuals; When individual fitness The current optimal value is shown in the following formula, where the random perturbation is the current position multiplied by 0.2 by the normal perturbation. 。 7. The photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network according to claim 1, characterized in that, The dataset is constructed by fusing and associating meteorological feature data and IMF components. An optimized Long Short-Term Memory (LSTM) network is trained using this dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted using the trained optimized LSM network, including: Each IMF component is fused with associated meteorological feature data, and input-output sample pairs are constructed according to a preset delay step size; The reconstructed sample pairs are randomly shuffled and then divided into training and test sets according to a preset ratio. Based on the optimized Long Short-Term Memory (LSTM) network, an independent optimized LSTM prediction model is constructed for each IMF component and residual term. The mean squared error is used as the loss function for model training. Using the optimized long short-term memory network models that have been trained, predictions are made based on the corresponding IMF components and residual terms to obtain the prediction results for each component and residual term. The predicted values ​​of all IMF components and residual terms are superimposed, and then restored to the original dimensions through inverse normalization to obtain the final photovoltaic power prediction result.

8. A photovoltaic power prediction system based on improved empirical mode decomposition and optimized long short-term memory network, characterized in that, include: The data acquisition module is used to collect and process historical photovoltaic power data and related meteorological data from photovoltaic power plants to obtain photovoltaic power time-series data and related meteorological characteristic data. An improved empirical mode decomposition module is used to perform empirical mode decomposition on photovoltaic power time series data based on endpoint extension and adaptive noise improvement, to obtain K IMF components and a residual term; The Long Short-Term Memory (LSTM) network parameter optimization module is used to optimize the number of hidden layer units, maximum training epochs, and initial learning rate of the LSM network with the goal of minimizing the global optimal root mean square error. This is achieved through a sparrow search algorithm improved by population initialization, adaptive weights, and vigilant perturbation, resulting in an optimized LSM network. A dataset is constructed by integrating correlated meteorological feature data and IMF components. An optimized long short-term memory network is trained on the dataset. Based on each IMF component and residual term, the final photovoltaic power result is predicted by the trained optimized long short-term memory network.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described in any one of claims 1-7.

10. An electronic device, characterized in that, Includes a memory and a processor: the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the photovoltaic power prediction method based on improved empirical mode decomposition and optimized long short-term memory network as described in any one of claims 1-7.

Citation Information

Cited By

  • A method and system for IGBT aging prediction

    CN122220807A