Short-term power load prediction method based on composite secondary variational mode decomposition (CIV-CBA) strategy

Abstract

The invention provides a short-term power load prediction method based on a composite secondary variational mode decomposition (CIV-CBA) strategy, belongs to the field of deep learning and signal processing, and aims to solve the problem that short-term power load prediction is crucial to efficient and stable operation of a power system, but the prediction precision is influenced by non-stationarity, nonlinearity and noise interference of power load data. A traditional single signal decomposition method is difficult to accurately capture dynamic characteristics, improper parameter setting easily causes excessive or insufficient decomposition, and meanwhile, gradient disappearance and explosion problems in neural network training also bring challenges. According to the method, self-adaptive noise complete ensemble empirical mode decomposition (CEEMDAN), variable mode decomposition (VMD), a convolutional neural network (CNN), a bidirectional gating recursion unit (BIGRU), an attention mechanism and an improved sparrow search algorithm (ISSA) are integrated, and by fusing multiple advanced technologies, the model can more accurately capture the intrinsic characteristics of power load data, so that the accuracy of the power load data is improved. Accuracy and stability of short-term power load prediction are remarkably improved, and powerful support is provided for efficient and stable operation of a power system.

Description

Technical Field

[0001] The present invention relates to short-term electric load forecasting technology, a power load forecasting model based on a composite quadratic variational mode decomposition (CIV-CBA) strategy, and belongs to the fields of deep learning and signal processing. Background Art

[0002] With the continuous growth of global energy demand and the complexity of power systems, short-term electric load forecasting has become an important technology to ensure the safe and efficient operation of power systems. Electric load forecasting not only helps in the rational planning of power systems, but also provides an important basis for the operation of power markets and the formulation of energy policies. However, electric load data usually exhibits complex characteristics such as non-linearity, non-stationarity, and high noise, which makes it difficult for existing traditional statistical models and machine learning models to accurately capture the dynamic change laws of the load, especially when dealing with high-noise and non-stationary data. In recent years, the rapid development of deep learning technology has provided new solutions for electric load forecasting. Deep learning models can automatically learn complex features in data and achieve accurate prediction of unknown data by simulating the working principle of the human brain neural network. In the field of electric load forecasting, deep learning technologies such as convolutional neural networks (CNNs), bidirectional gated recurrent units (BIGRUs), and long short-term memory networks (LSTMs) have been widely studied and applied. Although these models have achieved certain results under specific conditions, their performance is still greatly limited in the face of complex and variable electric load data. In addition, signal decomposition technologies have also been widely used in electric load forecasting, but single decomposition methods have problems of insufficient decomposition or over-decomposition caused by improper artificial parameter setting, thus affecting the accuracy and robustness of prediction. Therefore, how to effectively improve the accuracy and adaptability of short-term electric load forecasting remains an important problem to be solved in this field. Summary of the Invention

[0003] The present invention proposes a novel electric load forecasting model - an electric load forecasting model based on the composite improved variational mode decomposition (CIV-CBA) strategy. This model combines the advantages of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), variational mode decomposition (VMD), convolutional neural network (CNN), bidirectional gated recurrent unit (BIGRU), attention mechanism (Attention), and improved sparrow search algorithm (ISSA). Through the combination of quadratic decomposition and deep learning network, it can deeply explore the inherent dynamic characteristics of electric load data. Among them, in the model construction, CNN and BIGRU are innovatively combined to form a composite network structure with both spatial feature extraction and time series modeling capabilities. In the CNN layer, residual links and residual blocks are added to alleviate the gradient problem of deep networks and enhance the learning ability of complex features. The bidirectional processing mechanism of BIGRU comprehensively captures the context dependencies in the electric load sequence and improves the prediction accuracy. In addition, the attention mechanism is introduced to enable the model to focus on key data and identify complex dependencies between data. To optimize the selection of VMD parameters, the improved sparrow search algorithm is introduced, effectively reducing the volatility of load data.

[0004] The improvement of the improved sparrow search algorithm is as follows:

[0005] Compared with other swarm intelligence algorithms, the original SSA can usually find a better solution faster when solving complex optimization problems, showing higher convergence accuracy and speed, and having good robustness and stability. Despite these advantages, certain limitations hinder its performance. First, the random initialization of the population at the early stage hinders the realization of optimal ergodicity. Second, as the number of iterations increases, the diversity of the SSA population gradually decreases, resulting in the algorithm being prone to falling into local optimal solutions. To overcome these challenges, a series of improvement measures are proposed: initializing the population by combining chaotic maps (Logistics, Tent, Sine), using the minimum envelope entropy as the fitness function, updating the position of the discoverer using the butterfly optimization algorithm, and introducing the Levy flight strategy to update the position of the followers. These improvements aim to enhance the overall performance of the sparrow search algorithm.

[0006] 1. Initialize the population

[0007] Traditional SSA algorithms usually randomly generate the initial population in the search space, which may lead to different performances of the algorithm in multiple runs. If the initial population is unevenly distributed or of poor quality, the algorithm may require more iterations to find a better solution and may even fail to find the global optimal solution. To address this issue, a combined chaotic mapping (Logistics, Tent, Sine) is introduced to refine the initial population in SSA. This improvement enables the algorithm to be more effectively dispersed throughout the search space, making the initial state of the algorithm more diverse, thereby improving the global optimization performance and enhancing the robustness of the algorithm.

[0008] 2. Updating the position of the discoverer in the butterfly optimization algorithm

[0009] During the sparrow predation process, the ability of the discoverer to accurately locate the prey's position plays a decisive role in the foraging efficiency and action orientation of the entire group. However, the position update of this key role often faces the dilemma of falling into local optimal solutions due to individual differences and limitations of its own cognition. To overcome this defect, enhance the global search ability of the sparrow algorithm and prevent it from falling into local optimal solutions prematurely, the idea of the butterfly optimization algorithm is introduced to improve the position update stage of the discoverer.

[0010] 3. Updating the position of the followers using the Lévy flight strategy

[0011] The Lévy flight strategy simulates the random walk behavior of organisms during foraging or searching. Introducing the Lévy flight strategy in SSA allows the followers to more precisely track and capture the target prey based on the positions provided by the discoverer. By alternately using different step sizes, the followers can capture the prey faster. This strategy can not only expand the search space, increase the diversity of the population, but also effectively improve the situation where the algorithm falls into local optimal solutions, thereby enhancing the global exploration ability. Description of the drawings

[0012] Figure 1 It is a schematic diagram of the overall modeling framework of the CIV-CBA model;

[0013] Figure 2 It is a schematic diagram of the internal structure of the CNN-BIGRU-Attention model;

[0014] Figure 3 It is a schematic diagram of the ISSA-optimized VMD process. Detailed implementation manners

[0015] Taking the power load dataset as an example, the detailed implementation manners are as follows:

[0016] 1. Data preprocessing: First, collect the power load data for a recent period of time and preprocess it using CEEMDAN decomposition to extract the intrinsic features.

[0017] 2. Modal decomposition: Perform hierarchical clustering on the extracted data, dividing it into three types of signals: high-frequency, medium-frequency, and low-frequency. Then, further use the ISSA-VMD (Improved Sparrow Search Algorithm with Variational Mode Decomposition) algorithm to perform secondary modal decomposition on the high-frequency and medium-frequency components.

[0018] 3. Parameter optimization: Use the improved sparrow search algorithm (ISSA) to optimize the VMD parameters, find the best combination of the decomposition number K and the penalty factor α of VMD, so as to be able to more finely extract the potential patterns and features in the high-frequency components. This strategy can avoid the errors caused by manually setting parameters, thereby improving the overall performance of the model.

[0019] 4. Deep learning model training: For each IMF component, construct a CNN-BIGRU-Attention model for training. This model combines the advantages of the convolutional neural network (CNN), the bidirectional gated recurrent unit (BIGRU), and the attention mechanism (Attention). CNN is used to extract key features, BIGRU is used to capture the long-term dependencies of sequential data, and the Attention attention mechanism is used to highlight important information and suppress irrelevant information.

[0020] 5. Model evaluation: After completing the model training, sum the prediction results of each IMF component to obtain the final power load prediction result. According to the comparison between the prediction result and the actual load data, select five indicators: root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), coefficient of determination variance (R 2 )), mean absolute percentage error (MAPE) for comprehensive evaluation.

Claims

1. A power load forecasting method based on the Composite Improved Variational Mode Decomposition (CIV-CBA) strategy, characterized in that, The model includes the following steps: a) Perform data preprocessing on the collected power load data, and extract the intrinsic features of the power load data through Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN); b) Perform hierarchical clustering on the preprocessed data to divide it into three types of signals: high-frequency, medium-frequency, and low-frequency; c) Use the Improved Sparrow Search Algorithm (ISSA) to optimize the parameters of Variational Mode Decomposition (VMD), thereby determining the best combination of the decomposition number K and the penalty factor α, and achieving secondary mode decomposition of high-frequency and medium-frequency signals; d) Construct a deep learning model combining Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BIGRU), and Attention mechanism for each Intrinsic Mode Function (IMF) component for training; e) Add the prediction results of each IMF component obtained from training to obtain the final power load prediction result.

2. The power load forecasting model according to claim 1, characterized in that, The Improved Sparrow Search Algorithm (ISSA) includes the following improvements: a) Initialize the population through a combined chaotic map (Logistic, Tent, Sine) to improve the global optimization performance; b) Update the position of the discoverer using the Butterfly Optimization Algorithm to enhance the global search ability of the algorithm; c) Introduce the Lévy flight strategy to update the position of the followers, improving the search efficiency and robustness of the algorithm.

3. The power load forecasting model according to claim 1 or 2, characterized in that In the deep learning model, CNN is used to extract the spatial features of the power load data, BIGRU is used to capture the long-term dependencies of time series data, and the Attention mechanism is used to highlight important information and suppress irrelevant information.

4. The power load forecasting model according to claim 1, wherein In the model evaluation stage, five indicators, namely root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), coefficient of determination variance (R 2 ), and mean absolute percentage error (MAPE), are used to comprehensively evaluate the prediction results.

5. According to Claim 1, the model can adapt to the power load data in different seasons and time periods, and has good robustness and prediction accuracy.

6. The method according to claim 1, wherein The model can be trained based on historical load data to perform high-precision prediction of short-term power load, and can be widely applied to the fields of power system scheduling, load prediction, and smart grid, with good practical application value.

Images