Artificial intelligence-based power load forecasting and optimization method

Abstract

This invention discloses an artificial intelligence-based method for power load forecasting and optimization, specifically addressing the system stability problem caused by nonlinear resonance effects. It employs multi-scale decomposition and pattern recognition techniques to initially extract key features of power load signals and identify load patterns that may lead to system resonance. Focusing on two key indicators—frequency resonance intensity and time phase difference—a Bayesian network is used to comprehensively analyze the resonance risk of load patterns. A genetic algorithm is then combined to dynamically optimize the weight configuration of high-risk patterns, improving the model's sensitivity and adaptability to different resonance effects. Subsequently, the classified load patterns are input into a deep neural network for multi-pattern prediction fusion. A particle swarm optimization algorithm is used to adjust the model weights in real time, ensuring the prediction accuracy and operational stability of the system under complex load conditions, thus providing a strong guarantee for the reliability of the power system.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting and optimization, and more specifically, to a power load forecasting and optimization method based on artificial intelligence. Background Technology

[0002] Load forecasting is one of the core tasks in power system operation and management, and its accuracy directly affects the dispatch efficiency and operational safety of the power system. In power systems, the nonlinear dynamic behavior, seasonal fluctuations, and sudden changes in loads often increase the complexity of system operation. In particular, the stability of the power system is easily challenged when dealing with nonlinear resonance effects. With the continuous growth of electricity demand and the increasing complexity of load characteristics, traditional forecasting methods are proving inadequate in capturing and quantifying these dynamic changes, making it difficult to meet the requirements of modern power systems for high-precision load forecasting.

[0003] While existing technologies for power load forecasting incorporate artificial intelligence algorithms, such as support vector machines and neural networks, significant shortcomings remain. They typically rely solely on single time series of historical load data, failing to adequately consider the energy distribution of load patterns across different frequency components. This leads to a significant decrease in forecast accuracy when dealing with high-frequency resonance effects. Furthermore, they lack in-depth analysis of phase differences between load patterns, neglecting the impact of phase synchronization or phase shifts during critical periods on system stability. Moreover, most existing technologies employ fixed weight configurations, lacking dynamic adjustment mechanisms to flexibly optimize model parameters based on real-time risk assessment results. This results in insufficient predictive and system response capabilities in high-risk scenarios. These deficiencies make it difficult for existing methods to provide high-precision forecasts in complex power load environments, thus failing to effectively guarantee stable system operation.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide an artificial intelligence-based method for power load forecasting and optimization, aiming to address complex nonlinear dynamic behavior and potential resonance risks, and improve the prediction accuracy, stability, and operational efficiency of power systems. First, through multi-scale decomposition and feature extraction techniques, the power load signal is analyzed in depth to comprehensively capture its high-frequency and low-frequency variation characteristics, generating a multi-dimensional feature vector matrix with high accuracy and dynamic adaptability, providing a data foundation for subsequent nonlinear relationship detection. Second, by utilizing support vector machines combined with self-organizing mapping technology, implicit patterns in the load signal are successfully identified, especially key pattern combinations that may trigger system resonance, significantly enhancing the model's response capability and robustness under complex load conditions. Based on this, two key indicators—resonance intensity in the frequency direction and pattern phase difference in the time direction—are further used for comprehensive resonance risk assessment, and the weights of high-risk patterns are dynamically optimized using a genetic algorithm to ensure the model's flexibility and accuracy. Finally, the classified load patterns at different risk levels are input into a deep neural network for multi-pattern prediction fusion, and the pattern weights are adjusted in real time using a particle swarm optimization algorithm to ensure that the model provides optimal prediction results under different load variation conditions. The method of this invention not only optimizes the accuracy of the power load forecasting model, but also significantly improves the operational stability and security of the power system in complex environments, providing strong technical support for the long-term reliable operation of the system and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1 uses wavelet transform to decompose historical power load data at multiple scales and extracts load pattern features at each frequency level; it then uses Hilbert-Huang transform to obtain the instantaneous amplitude and instantaneous frequency of the frequency components and generates a multidimensional feature vector.

[0008] S2 inputs the multidimensional feature vectors into the support vector machine and identifies the nonlinear modes in the load through the kernel method; then, it uses self-organizing maps to perform cluster analysis on the modes and identify potential combinations of resonant modes.

[0009] S3 combines two indicators—resonance intensity in the frequency direction and mode phase difference in the time direction—and uses a Bayesian network to comprehensively analyze resonance risk, classifying load patterns into high, medium, and low risk levels; and uses a genetic algorithm to dynamically adjust the weights of load patterns identified as high-risk.

[0010] S4. Based on the risk classification results, the classified patterns are input into a deep neural network for multi-pattern prediction fusion, and the pattern weights are adjusted in real time through particle swarm optimization to optimize the model prediction accuracy.

[0011] In a preferred embodiment, step S1 includes the following:

[0012] S1.1, Preprocess historical power load data;

[0013] S1.2 Input the preprocessed power load data into the wavelet transform model, and use the multi-scale decomposition method to decompose the original load data into frequency components of different scales, generating multiple frequency band signals, each corresponding to the load pattern at different time scales.

[0014] S1.3, taking the wavelet decomposition signals of each frequency band as input, and using Hilbert-Huang transform to analyze each frequency band signal; first, applying empirical mode decomposition to decompose the signal into intrinsic mode functions; then, performing Hilbert transform on each IMF to obtain the instantaneous amplitude and instantaneous frequency of the signal.

[0015] S1.4 combines the instantaneous amplitude and instantaneous frequency of each frequency band with the original wavelet coefficient matrix to construct a multidimensional feature vector.

[0016] S1.5 stores the constructed multidimensional feature vector matrix and associates it with the corresponding timestamp and power load scenario.

[0017] In a preferred embodiment, step S2 includes the following:

[0018] S2.1, take the stored multidimensional feature vector matrix as input data and perform preprocessing.

[0019] S2.2, the preprocessed multidimensional feature vectors are input into the SVM model for training, optimizing the model's decision boundary so that it can distinguish different load patterns.

[0020] S2.3. After the SVM model is trained, the multidimensional feature vector is input into the trained model again to perform nonlinear pattern recognition.

[0021] S2.4. Use the nonlinear patterns identified by SVM as input to perform self-organizing map clustering analysis;

[0022] S2.5 combines the cluster analysis pattern features with the identified resonance patterns and stores them, associating them with their corresponding timestamps and power load scenarios.

[0023] In a preferred embodiment, step S3 includes the following:

[0024] S3.1 Based on the nonlinear modes identified in step S2, the resonance intensity of each load mode in different frequency ranges is first extracted.

[0025] S3.2, Based on the identification of resonance intensity, calculate the mode phase difference in the time direction;

[0026] S3.3 uses the phase difference matrix and the phase difference matrix as input variables to construct a Bayesian network model to determine the risk level of the nonlinear resonance effect.

[0027] Based on the posterior probability distribution, the resonance risk level of the system is determined; load patterns with posterior probabilities greater than or equal to the high-risk threshold are classified as high-risk.

[0028] Load patterns with posterior probabilities not less than the high-risk threshold and greater than or equal to the low-risk threshold are classified as medium-risk.

[0029] Patterns with posterior probabilities less than the low-risk threshold are classified as low-risk.

[0030] S3.4 Based on the risk analysis results of the Bayesian network, the weights of the identified high-risk load pattern combinations are dynamically adjusted; the weights of the high-risk patterns are preferentially optimized using a genetic algorithm.

[0031] S3.5 stores the optimized high-risk load pattern weights and associates them with the corresponding timestamps and power load scenarios.

[0032] In a preferred embodiment, the complete steps for obtaining the resonant intensity of each load mode in different frequency ranges are as follows:

[0033] S3.1.1 First, the preprocessed power load signal x(t) is converted to the frequency domain, and a Fourier Transform (FT) is used to obtain the frequency domain signal X(f), where f represents the frequency; the frequency range F is constructed as multiple discrete intervals F = {f1, f2, ..., f...} n}, where f i This represents the i-th frequency interval; by performing a Fourier transform on the power load signal, the spectral signal X(f) of each frequency interval is obtained. i ):

[0034] S3.1.2, after obtaining the spectral signal for each frequency interval, calculate the instantaneous energy density for each frequency interval; construct the instantaneous energy density E(f i Let ,t) be the square form of the signal energy at time t; it is calculated using a nonlinear transformation of energy density, using the following formula: Among them, α, β, and γ are nonlinear transformation parameters used to adjust the calculation method of instantaneous energy density in order to capture the dynamic changes of the signal in different frequency ranges.

[0035] S3.1.3, after calculating the instantaneous energy density for each frequency range, extract the resonance factor R(f) within the corresponding frequency range. i ): Where, θ(f) i ,t) represents the frequency range f i The instantaneous phase angle is obtained from the time-domain signal through Hilbert transform.

[0036] S3.1.4, for each frequency range, the resonance factor, representing the original resonance intensity of the corresponding frequency range, is normalized; the normalized resonance intensity is denoted as S(f i ).

[0037] S3.1.5 Finally, the normalized resonance intensities of each frequency range are organized into a resonance intensity matrix D(M), where each row of the matrix corresponds to a frequency range and each column corresponds to the resonance intensity of different time windows.

[0038] In a preferred embodiment, the complete steps for obtaining the mode phase difference in the time direction are as follows:

[0039] S3.2.1, First, the instantaneous phase information is extracted from the power load signal x(t) using the Hilbert transform; for a given power load signal, its corresponding analytic signal z(t) is obtained through the Hilbert transform, and the analytic signal is defined as: in, The Hilbert transform of x(t) is represented by j, where j is the imaginary unit; the instantaneous phase θ(t) is calculated from the phase angle of the analytic signal: θ(t) = θ(t) represents the phase change of the load signal at each moment, laying the foundation for subsequent phase difference analysis.

[0040] S3.2.2, after obtaining the instantaneous phase, calculate the phase synchronization index (PLV) to measure the degree of phase synchronization between the two load modes x1(t) and x2(t); the formula for calculating the phase synchronization index PLV is: Where N is the number of sampling points within the time window, θ1(t k ) and θ2(t k ) represent the two modes at time t k The instantaneous phase.

[0041] S3.2.3, After obtaining the phase synchronization index, calculate the phase difference metric (PLI) to describe the directionality of the phase offset; the formula for calculating the phase difference metric (PLI) is: Where sign(·) represents the sign function, which is used to preserve the directionality of the phase shift, and λ is a nonlinear extension parameter used to adjust the sensitivity of the phase difference measurement.

[0042] S3.2.4 Integrate the phase difference measurements within each time window to calculate the overall phase difference intensity D(T), as shown in the formula: Where T represents the total time interval [t0, t1], t c ρ is the time center point, and ρ is the time decay factor, used to adjust the influence of different time periods on the overall phase difference intensity.

[0043] S3.2.5 Finally, the overall phase difference intensity calculated at each time window will be organized into a phase difference matrix S(F), where each row of the matrix corresponds to a load mode combination and each column corresponds to the phase difference intensity at different time periods.

[0044] In a preferred embodiment, step S4 includes the following:

[0045] S4.1 inputs the load patterns, which have been classified into high, medium, and low risk levels, into the deep neural network and performs preprocessing before input.

[0046] S4.2, based on the preprocessed load pattern data, uses a deep neural network to perform multi-pattern prediction fusion for high, medium and low risk levels.

[0047] S4.3, based on the trained deep neural network, uses the particle swarm optimization algorithm to adjust the weights of the load pattern in real time.

[0048] S4.4 After completing the weight adjustment, the deep neural network outputs the fused prediction result.

[0049] The technical effects and advantages of the artificial intelligence-based power load forecasting and optimization method of this invention are as follows:

[0050] 1. By performing multi-scale decomposition and feature extraction on the power load data in step S1, not only can the high-frequency and low-frequency variation characteristics in the load signal be comprehensively captured, but key feature points strongly correlated with power load patterns can also be effectively identified. The multi-dimensional feature vector matrix constructed in this process has good time-frequency accuracy and dynamic adaptability, providing a high-quality data foundation for nonlinear relationship detection and complex load pattern identification. With the help of an efficient data management system and dynamic update mechanism, the feature vector matrix can be continuously optimized with the introduction of new data, enabling the power load forecasting model to have stronger prediction accuracy and adaptability when facing future load changes, ultimately improving the stability and operating efficiency of the power system.

[0051] 2. By utilizing Support Vector Machines (SVM) for nonlinear pattern recognition in step S2 and combining it with Self-Organizing Maps (SOM) for cluster analysis, the implicit nonlinear patterns in the power load signals were successfully identified, especially those key pattern combinations that could trigger system resonance. This process effectively improved the ability to identify complex load patterns, enabling the prediction model to more accurately reflect the dynamic behavior of the power system. Identifying these nonlinear structures not only provides accurate data support for subsequent resonance risk assessment and handling but also improves the robustness and stability of the overall prediction model, ultimately enhancing the power system's response capability and optimization performance in the face of nonlinear challenges.

[0052] 3. The beneficial effect of step S3 lies in achieving a comprehensive and accurate identification of nonlinear resonance effects in power systems through the combined analysis of two indicators: resonance intensity in the frequency direction and mode phase difference in the time direction. First, frequency-direction resonance intensity analysis effectively detects the energy accumulation of power load signals in specific frequency ranges, revealing potential resonance risks at these frequencies, particularly high-frequency resonance. This analysis allows the system to provide early warnings of high-energy points that may lead to instability, enabling preventative measures to be taken in advance. Second, time-direction mode phase difference analysis captures the phase synchronization and offset between different load modes in the time domain, accurately identifying the dynamic behavior of the system during critical periods, especially identifying phase offsets and coupling effects between different modes that may lead to system misalignment or resonance. This combined analysis not only expands the depth and breadth of risk identification but also achieves a comprehensive assessment of complex nonlinear risks through synergy, avoiding omissions and misjudgments that may arise from a single analytical dimension. Meanwhile, by dynamically adjusting the weights of high-risk patterns through genetic algorithms, the model can respond to potential system risks more flexibly and accurately, thereby significantly improving the adaptability, robustness, and overall prediction accuracy of the power load forecasting model, and providing a reliable guarantee for the stable operation of the power system.

[0053] 4. By classifying load patterns into high, medium, and low risk levels in step S4 and inputting the data into a deep neural network for multi-mode prediction fusion, the model's ability to identify and predict complex power load patterns is significantly improved. Prioritizing high-risk patterns ensures the system's response speed and accuracy in the face of potential major risks, while moderate adjustment of medium- and low-risk patterns maintains the overall balance of the model and avoids excessive resource allocation. The introduction of particle swarm optimization allows the model to dynamically adjust mode weights based on real-time data during actual operation, ensuring that the contribution of each mode to the prediction results remains optimal across different time periods. This innovative approach not only improves the accuracy and robustness of the prediction model but also enhances the stability and security of the power system in the face of complex load changes, providing strong support for the reliable operation of the system. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the power load prediction and optimization method based on artificial intelligence according to the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] Figure 1 This invention presents an artificial intelligence-based method for power load forecasting and optimization, comprising:

[0058] S1 uses wavelet transform to decompose historical power load data at multiple scales and extracts load pattern features at each frequency level; it then uses Hilbert-Huang transform to obtain the instantaneous amplitude and instantaneous frequency of the frequency components and generates a multidimensional feature vector.

[0059] S2 inputs the multidimensional feature vectors into the support vector machine and identifies the nonlinear modes in the load through the kernel method; then, it uses self-organizing maps to perform cluster analysis on the modes and identify potential combinations of resonant modes.

[0060] S3 combines two indicators—resonance intensity in the frequency direction and mode phase difference in the time direction—and uses a Bayesian network to comprehensively analyze resonance risk, classifying load patterns into high, medium, and low risk levels; and uses a genetic algorithm to dynamically adjust the weights of load patterns identified as high-risk.

[0061] S4. Based on the risk classification results, the classified patterns are input into a deep neural network for multi-pattern prediction fusion, and the pattern weights are adjusted in real time through particle swarm optimization to optimize the model prediction accuracy.

[0062] In modern power systems, load forecasting and optimization are core tasks for ensuring system stability and efficient operation. With the continuous increase in electricity demand and the increasing complexity of load patterns, traditional forecasting methods are struggling to cope with the frequent nonlinear dynamic behaviors and potential resonance risks in power systems. These risks can not only reduce system operating efficiency but also trigger serious stability problems, affecting the reliability of the entire power system. To overcome this challenge, this invention proposes a power load forecasting and optimization method based on multi-scale analysis, nonlinear pattern recognition, and comprehensive risk assessment. By performing multi-scale decomposition of historical power load data using a combination of wavelet transform and Hilbert-Huang transform, load pattern features at each frequency level are extracted, generating multi-dimensional feature vectors to capture complex dynamic behaviors in power signals. Based on this, support vector machines and self-organizing maps are used to perform nonlinear pattern recognition and cluster analysis on these feature vectors, identifying key pattern combinations that may trigger resonance. Subsequently, a Bayesian network is used to comprehensively analyze two indicators: resonance intensity in the frequency direction and pattern phase difference in the time direction, classifying and assessing the potential resonance risks of the system. Finally, a genetic algorithm is used to dynamically adjust the weights of the identified high-risk load patterns. Finally, based on the risk classification results, these classified patterns are input into a deep neural network for multi-mode prediction fusion. Particle swarm optimization is then used to adjust the pattern weights in real time, ensuring the accuracy of the prediction model and the operational stability of the power system. This invention's method, through an innovative combination of technologies, achieves accurate prediction and optimization of power load, significantly improving the power system's ability to cope with complex load changes and potential risks, and ensuring the system's reliability and efficiency in complex operating environments.

[0063] In the process of power load forecasting and optimization, accurately extracting and analyzing key features from historical load data is fundamental to ensuring the high accuracy and adaptability of the forecasting model. In complex power systems, load changes at different time scales contain rich information. The distribution and interrelationships of this information in the frequency and time domains directly affect the identification of load patterns and the detection of nonlinear relationships. By performing multi-scale decomposition on historical power load data and extracting its multi-dimensional feature vectors, key data support can be provided for subsequent nonlinear analysis and forecasting models, thereby improving the overall forecasting accuracy and the model's responsiveness.

[0064] Step S1 includes the following:

[0065] S1.1, Preprocessing of Historical Power Load Data. During data preprocessing, firstly, based on the specific needs of power load forecasting, the time range and granularity of the historical data to be processed are determined, and data relevant to the current analysis objective is selected, while irrelevant historical data is excluded. Then, within the selected data range, statistical methods such as the Three Sigma (3σ) criterion or box plots are used to detect outliers, removing obviously erroneous data points or replacing them with linear interpolation algorithms. Load data from peak hours, holidays, or extreme weather periods are prioritized for processing, employing stricter noise filtering and high-precision interpolation algorithms to ensure that this valuable data receives more detailed processing. Afterwards, the filtered and processed data undergoes Min-Max normalization or Z-score standardization to uniformly adjust the data to the same scale range, eliminating differences between different time periods or units, and ensuring consistent model sensitivity to different features. Finally, the preprocessed data is stored hierarchically, managed according to time range, load characteristics, and data quality, allowing for flexible retrieval in subsequent analyses and dynamic adjustment of data priority to meet the needs of different analytical tasks.

[0066] In step S1.2, the preprocessed power load data is input into a wavelet transform model, and a suitable mother wavelet function, such as the Daubechies wavelet (db4), is selected. Using a multi-scale decomposition method, the original load data is decomposed into frequency components at different scales, generating multiple frequency bands corresponding to load patterns at different time scales. The load pattern features of each frequency band are extracted using the wavelet coefficient matrix, recording high-frequency and low-frequency information at each scale and preserving key load change characteristics.

[0067] Key load variation characteristics can be defined as the dynamic behavior of high-frequency and low-frequency components of the power load signal at different frequency levels after multi-scale decomposition. Specifically, this includes drastic fluctuations in instantaneous amplitude, abnormal frequency shifts, and characteristic points strongly correlated with load patterns in specific time periods. By analyzing the energy distribution, time-domain and frequency-domain cross-relationships, and phase consistency of these high-frequency and low-frequency information at different time periods, features that significantly influence the overall load pattern can be extracted. These features include, but are not limited to, energy accumulation effects at specific frequencies, load spikes caused by external factors, trends in periodic fluctuations, and phase differences reflecting the system's load response speed. Key load variation characteristics defined in this way can comprehensively reflect the power system's sensitivity and responsiveness to various load changes at different time scales, providing a more accurate and practically guiding data foundation for subsequent nonlinear relationship detection and load forecasting.

[0068] In step S1.3, the signals from each frequency band after wavelet decomposition are used as input, and the Hilbert-Huang Transform (HHT) is applied for further analysis of each frequency band. First, Empirical Mode Decomposition (EMD) is applied to decompose the signal into Intrinsic Mode Functions (IMFs), ensuring that each IMF satisfies the definition of instantaneous frequency. Then, a Hilbert Transform is performed on each IMF to obtain the instantaneous amplitude and instantaneous frequency of the signal, forming a dynamic characteristic description of the frequency components.

[0069] In step S1.4, the instantaneous amplitude and frequency of each frequency band are combined with the original wavelet coefficient matrix to construct a multidimensional feature vector. This feature vector not only contains the frequency characteristics of the load signal but also encompasses its variation trend over time. By classifying and standardizing the instantaneous features of different frequency components, the final multidimensional feature vector matrix is ​​generated, which serves as the input data for subsequent nonlinear relationship detection.

[0070] Step S1.5 stores the constructed multidimensional feature vector matrix and associates it with corresponding timestamps and power load scenarios. By building an efficient data management system, it ensures that these feature vectors can be quickly extracted and retrieved during nonlinear relationship detection and pattern recognition. To cope with future changes in power load, a dynamic update mechanism is set up, enabling the feature vector matrix to be updated and optimized in real time as new data is introduced.

[0071] The above steps ensure that the features extracted from historical power load data have high time-frequency accuracy and flexibility, providing a solid data foundation for the accurate detection of nonlinear relationships.

[0072] By performing multi-scale decomposition and feature extraction on the power load data in step S1, not only can the high-frequency and low-frequency variation characteristics of the load signal be comprehensively captured, but key feature points strongly correlated with power load patterns can also be effectively identified. The multi-dimensional feature vector matrix constructed in this process has good time-frequency accuracy and dynamic adaptability, providing a high-quality data foundation for nonlinear relationship detection and complex load pattern identification. With the help of an efficient data management system and dynamic update mechanism, the feature vector matrix can be continuously optimized with the introduction of new data, enabling the power load forecasting model to have stronger prediction accuracy and adaptability when facing future load changes, ultimately improving the stability and operating efficiency of the power system.

[0073] In power load forecasting and optimization, the underlying nonlinear patterns in power load signals have a significant impact on system stability and forecast accuracy. With increasing power system complexity, traditional linear analysis methods struggle to capture these intricate nonlinear relationships, leading to biased forecasts and reduced optimization effectiveness. By combining Support Vector Machines (SVM) with kernel methods for nonlinear pattern recognition and utilizing Self-Organizing Maps (SOM) for pattern clustering analysis, the complex dynamic characteristics of load signals can be accurately revealed, particularly those patterns closely related to system resonance risk. This provides a crucial foundation for subsequent resonance effect prediction and mitigation strategies.

[0074] Step S2 includes the following:

[0075] Step S2.1 uses the stored multidimensional feature vector matrix as input data. First, the feature vectors are preprocessed to ensure they are suitable for training and prediction using the Support Vector Machine (SVM) model. Specific preprocessing includes standardizing the feature vectors to eliminate scale differences between different features, and using PCA (Principal Component Analysis) to reduce dimensionality, retaining the main information in the feature vectors, reducing data redundancy, and improving computational efficiency.

[0076] S2.2. Based on the complexity and nonlinear characteristics of power load, a suitable kernel method is selected. Kernel functions such as the Radial Basis Function (RBF) kernel or the Polynomial Kernel, which can capture nonlinear relationships, are used to construct the SVM model. Cross-validation is used to determine the optimal kernel parameters and regularization parameters to ensure that the model can accurately identify nonlinear patterns in the load data. The preprocessed multidimensional feature vectors are then input into the SVM model for training, optimizing the model's decision boundary to enable it to distinguish different load patterns.

[0077] S2.3. After the SVM model is trained, the multidimensional feature vectors are input again into the trained model for nonlinear pattern recognition. Based on the decision boundary formed during training, the model classifies the input data and identifies different pattern characteristics in the power load, especially those load patterns with nonlinear relationships. By identifying these patterns, the implicit nonlinear structure in the power load signal can be revealed, providing a foundation for subsequent pattern analysis.

[0078] In nonlinear pattern recognition, preprocessed multidimensional feature vectors are input into a pre-trained Support Vector Machine (SVM) model. The model classifies the input data one by one based on the optimal decision boundary constructed during training. During classification, the model maps the feature vectors using kernel methods, transforming the originally complex high-dimensional data into a higher-dimensional feature space, thus making it easier to separate different patterns in the new space. Specifically, SVM utilizes its high-precision boundary partitioning capability to identify different power load pattern characteristics hidden in the input data, especially nonlinear patterns that are difficult to capture using linear methods. These nonlinear patterns typically reflect complex dynamic behaviors in power load signals, such as sudden load changes, nonlinear enhancement of periodic fluctuations, and nonlinear resonance caused by coupling with external factors. Accurate identification of these patterns reveals potential hidden nonlinear structures in the power system. These structures have important reference value in subsequent pattern analysis and resonance risk assessment, providing more detailed and accurate data support for power load forecasting and optimization.

[0079] S2.4. Using the nonlinear patterns identified by SVM as input, perform Self-Organizing Map (SOM) clustering analysis. The SOM model uses unsupervised learning to perform topological mapping on pattern features, clustering similar patterns into the same class. By analyzing the clustering results, identify pattern combinations with potential resonance effects, which may induce nonlinear resonance in the power system under specific conditions.

[0080] S2.5 stores the clustered pattern features and the identified resonance pattern combinations, associating them with their corresponding timestamps and power load scenarios for quick retrieval in subsequent resonance risk prediction. Simultaneously, a dynamic update mechanism for pattern features is established, updating the definition and classification of resonance pattern combinations in real time as new data is introduced and the model is retrained to adapt to dynamic changes in power load.

[0081] Through the above steps, we can ensure that the nonlinear modes in the power load are accurately identified from the multidimensional feature vectors, and reveal potential resonant mode combinations through cluster analysis, providing solid data support for subsequent resonant risk prediction and response strategies.

[0082] By utilizing Support Vector Machines (SVM) for nonlinear pattern recognition in step S2 and combining it with Self-Organizing Maps (SOM) for cluster analysis, implicit nonlinear patterns in power load signals were successfully identified, particularly key pattern combinations that could trigger system resonance. This process effectively improved the ability to identify complex load patterns, enabling the prediction model to more accurately reflect the dynamic behavior of the power system. Identifying these nonlinear structures not only provides precise data support for subsequent resonance risk assessment and handling but also improves the robustness and stability of the overall prediction model, ultimately enhancing the power system's responsiveness and optimization performance in the face of nonlinear challenges.

[0083] Step S3 includes the following:

[0084] In power load forecasting and optimization, identifying and quantifying the resonance intensity of load patterns across different frequency ranges is crucial for understanding the dynamic behavior and potential risks of the power system. Nonlinear resonance effects in power systems often concentrate within specific frequency intervals, potentially leading to abnormal system fluctuations and impacting power supply stability. Detailed analysis of the resonance intensity of various load patterns across different frequency ranges can reveal key dynamic characteristics in load signals, providing crucial data support for the stable operation of the power system and offering strong evidence for optimizing forecasting models and providing early warnings of resonance risks.

[0085] The rationale for combining the frequency-direction resonance intensity and the time-direction mode phase difference in risk identification is that these two indicators can comprehensively capture the nonlinear dynamic behavior of the power system, thereby achieving accurate identification of potential resonance risks. Frequency-direction resonance intensity, by analyzing the energy accumulation of power load signals in different frequency ranges, can reveal the resonance effects that may occur in the system at specific frequencies, and is particularly suitable for identifying high-frequency resonance problems caused by external disturbances or the system's own instability. Meanwhile, the time-direction mode phase difference, by measuring the phase synchronization and phase shift between different load modes, can effectively identify the coordinated or misaligned state of the system within a specific time period. Especially when there are complex coupling relationships between load modes, this indicator can reflect the dynamic instability and potential nonlinear resonance risks of the system in the time domain. The combined analysis of these two indicators not only comprehensively captures the dynamic behavior of the system from both the frequency and time domains, but also, through their synergistic effect, more accurately identifies complex risk modes that are difficult to detect with a single indicator, thus providing stronger support for system stability analysis and risk early warning.

[0086] S3.1, Based on the nonlinear modes identified in step S2, the resonance intensity of each load mode is first extracted within different frequency ranges. The resonance intensity is calculated based on the energy density variation of each frequency component within a specific time window. The load signal is converted to the frequency domain using Fourier transform, and the energy distribution in different frequency bands is analyzed to identify high-energy points where resonance effects may occur. These high-energy points are then quantitatively analyzed to generate resonance intensity indices. Specific calculations include the time-varying characteristics of energy density and its average value and standard deviation over multiple time periods to accurately reflect the resonance risk at different frequencies.

[0087] Complete steps to obtain the resonant intensity of each load mode in different frequency ranges:

[0088] S3.1.1 First, the preprocessed power load signal x(t) is converted to the frequency domain, and a Fourier Transform (FT) is used to obtain the frequency domain signal X(f), where f represents the frequency. The frequency range F is constructed as multiple discrete intervals F = {f1, f2, ..., f...} n}, where f i Let X(f) represent the i-th frequency interval. By performing a Fourier transform on the power load signal, the spectral signal X(f) for each frequency interval is obtained. i ):

[0089] S3.1.2, after obtaining the spectral signal for each frequency interval, calculate the instantaneous energy density for each frequency interval. Construct the instantaneous energy density E(f i Let ,t) represent the squared form of the signal energy at time t. Considering the nonlinear characteristics of the power load signal, a nonlinear transformation of the energy density is used for calculation, as shown in the following formula: Where α, β, and γ are nonlinear transformation parameters used to adjust the calculation method of instantaneous energy density in order to capture the dynamic changes of the signal in different frequency ranges. Here, X(f i ) represents the frequency range f i The corresponding signal amplitude.

[0090] S3.1.3, after calculating the instantaneous energy density for each frequency range, further extract the resonance factor R(f) within the corresponding frequency range. i The resonance factor reflects the degree of signal energy concentration within a given frequency range and its impact on load patterns. The formula for calculating the resonance factor is: Where, θ(f) i ,t) represents the frequency range f iThe instantaneous phase angle is obtained from the time-domain signal through Hilbert transform. The integration range [t0, t1] is a specific time period, chosen to focus on the critical time window where the resonance effect occurs. By introducing the phase angle, the resonance factor can better reflect the resonance characteristics of the signal within this frequency range.

[0091] S3.1.4, for each frequency range, the original resonance intensity of the corresponding frequency range is represented by a resonance factor. To unify the representation of resonance intensity across different frequency ranges, the resonance factor needs to be normalized. The normalized resonance intensity S(f) i The formula for calculating ) is: Here, δ is a normalization parameter used to adjust the relative weights between resonance factors in different frequency ranges, ensuring that the normalized resonance intensity is within the range of 0 to 1. This eliminates energy differences between different frequency ranges, facilitating unified processing in subsequent analyses.

[0092] In step S3.1.5, the normalized resonance intensities for each frequency range are finally organized into a resonance intensity matrix D(M). Each row of the matrix corresponds to a frequency range, and each column corresponds to the resonance intensity for different time windows. This matrix constitutes a comprehensive description of the resonance effect of the power load signal in different frequency ranges, providing basic data for subsequent Bayesian network risk analysis and dynamic adjustment of load patterns.

[0093] Through the above steps, the resonance intensity of power load patterns can be accurately obtained within different frequency ranges, thereby providing reliable data support and analysis basis for the stable operation of the power system.

[0094] By accurately calculating and analyzing the resonance intensity of various load modes within different frequency ranges, potential resonance effects in the power system and the dynamic instability they may induce can be effectively identified. This process not only helps in understanding the complex dynamic behavior of power load signals but also provides high-precision input data for subsequent risk assessment and prediction models through the normalized resonance intensity matrix. Ultimately, using this resonance intensity data, the adaptability and responsiveness of power load prediction models to nonlinear resonance effects can be significantly improved, ensuring that the system maintains a stable operating state in complex and ever-changing power environments and enhancing the overall reliability of prediction and optimization.

[0095] In power load forecasting and optimization, the phase synchronization and phase difference between load patterns are crucial factors affecting system stability. Phase shifts between different patterns in the power load signal, especially phase synchronization or misalignment within specific time periods, can lead to nonlinear resonance effects in the system, subsequently causing abnormal fluctuations or faults. Therefore, accurately acquiring and quantifying the phase differences of power load patterns in the time direction is of great significance for identifying potential system risks, improving the adaptability of forecasting models, and optimizing system operation. Precise time-domain analysis of the phase differences of load patterns allows for a deeper understanding of the system's dynamic behavior, providing strong support for the safe and stable operation of the power system.

[0096] S3.2, based on the identification of resonance intensity, further calculate the mode phase difference in the time direction. Mode phase difference refers to the phase synchronization or difference between different load modes in the time domain, with particular attention paid to mode combinations that exhibit phase shifts at critical moments. Hilbert transform is used to obtain the instantaneous phase information of the load signal, and the phase difference between different modes is quantitatively assessed using indicators such as the Phase Locking Value (PLV) or the Phase Lag Index (PLI). This indicator reflects the cooperative or antagonistic relationship of load modes at different time periods and is an important basis for measuring the stability of the system in the time dimension.

[0097] Complete steps to obtain mode phase differences in the time direction:

[0098] S3.2.1 First, the instantaneous phase information is extracted from the power load signal x(t) using the Hilbert Transform. For a given power load signal, its corresponding analytic signal z(t) is obtained through the Hilbert Transform. The analytic signal is defined as: in, Let x(t) represent the Hilbert transform, where j is the imaginary unit. The instantaneous phase θ(t) is calculated from the phase angle of the analytic signal. θ(t) represents the phase change of the load signal at each moment, laying the foundation for subsequent phase difference analysis.

[0099] S3.2.2, after obtaining the instantaneous phase, consider the phase synchronization between different load modes. Calculate the Phase Locking Value (PLV) to measure the degree of phase synchronization between two load modes x1(t) and x2(t). The formula for calculating the Phase Locking Value (PLV) is: Where N is the number of sampling points within the time window, θ1(t k ) and θ2(tk ) represent the two modes at time t k The instantaneous phase. Phase synchronization refers to a value ranging from 0 to 1. The closer the value is to 1, the higher the phase synchronization between the two modes, and vice versa.

[0100] S3.2.3, After obtaining the phase synchronization index, to further quantify the phase difference between different modes, the Phase Lag Index (PLI) is calculated to describe the directionality of the phase shift. The formula for calculating the Phase Lag Index (PLI) is as follows: Here, sign(·) represents the sign function, used to preserve the directionality of the phase shift, and λ is a nonlinear extension parameter used to adjust the sensitivity of the phase difference metric. The phase difference metric reflects the frequency of phase shifts between modes and their degree of clustering in a specific direction.

[0101] S3.2.4, In order to quantify the overall phase difference of the mode in the time direction, the phase difference metric values ​​within each time window are integrated to calculate the overall phase difference intensity D(T), as shown in the formula: Where T represents the total time interval [t0, t1], t c Let be the time center point, and ρ be the time decay factor, used to adjust the influence of different time periods on the overall phase difference intensity. This is achieved through an exponential decay function. This results in phase difference measures closer to the center time contributing more to the overall phase difference intensity, while those farther from the center time contribute less.

[0102] S3.2.5 Finally, the overall phase difference intensity calculated over each time window is organized into a phase difference matrix S(F). Each row of the matrix corresponds to a load pattern combination, and each column corresponds to the phase difference intensity over different time periods. The phase difference matrix provides a comprehensive description of the phase difference of different load patterns in the time direction, providing crucial data support for subsequent Bayesian network analysis and dynamic adjustment of load pattern weights.

[0103] By following the steps above, the phase difference of the power load pattern in the time direction can be accurately obtained, and the phase synchronization and offset between different patterns can be quantified, providing a scientific basis for the stability analysis and prediction of the system.

[0104] By analyzing the phase differences of power load patterns over time, we can effectively identify the phase synchronization and offset between different patterns, revealing potential dynamic instability factors in the power system. This process not only helps identify high-risk pattern combinations that are prone to resonance effects, but also provides accurate data support for subsequent Bayesian network analysis and load pattern weight adjustment. By constructing a phase difference matrix, we can comprehensively reflect the phase dynamics of the system at different time periods, providing more detailed and accurate input data for the prediction model. Ultimately, this improves the model's response to nonlinear resonance effects and enhances the operational stability and safety of the power system under complex load conditions.

[0105] In power systems, nonlinear resonance effects often pose a hidden risk of system instability and failure. With the increasing complexity and dynamic changes of power loads, traditional linear analysis methods struggle to fully capture the potential nonlinear correlations and risk factors within the system. Bayesian networks, as a powerful probabilistic graphical model, can achieve comprehensive risk analysis of complex systems by modeling the causal relationships between different variables. By constructing a Bayesian network model using the phase difference matrix and resonance intensity matrix as input variables, we can deeply analyze the dynamic characteristics between power load patterns from both time and frequency dimensions, and quantify and provide early warnings of risks that may trigger system resonance. This analytical method provides a new perspective and tool for power system operation optimization and risk management, helping to improve system stability and security.

[0106] Section S3.3 uses the phase difference matrix and phase difference matrix as input variables to construct a Bayesian Network model for comprehensive risk analysis of nonlinear resonance effects. The Bayesian Network calculates the probability of different resonance mode combinations and their impact on system stability by constructing a probability relationship graph between variables and combining historical and real-time data. Based on the resonance intensity in the frequency direction and the mode phase difference in the time direction, the Bayesian Network can dynamically adjust the conditional probability distribution of each node, output the risk level of each resonance mode combination, identify the potential resonance effects with the greatest impact on the power system, and when the posterior probability exceeds the corresponding threshold, it can be determined that the current load mode combination has a significant resonance risk, outputting a high-risk signal.

[0107] In section S3.3.1, the phase difference matrix and resonance intensity matrix are used as input variables for constructing the Bayesian network model. These two matrices reflect the dynamic characteristics of different load mode combinations in the time and frequency directions, respectively. For each load mode combination, its phase difference D within a specific time period is extracted. i and resonance intensity S j The values ​​of these two matrices serve as the input node values. To ensure the effectiveness of the Bayesian network modeling, the data in these two matrices needs to be preprocessed.

[0108] Data discretization: Since Bayesian networks typically process discrete data, it is necessary to discretize the continuous values ​​of phase difference and resonance intensity. This can be done by dividing the data into several intervals, such as low, medium, and high-risk intervals. The number of discretization intervals and the method of division can be adjusted according to actual needs and data distribution.

[0109] Variable Combinations: For each load pattern combination, the input variable V is formed. k Its form is V k =(D i ,S j The joint variables contain the dynamic characteristics of the load pattern in both time and frequency directions and are the input nodes of the Bayesian network.

[0110] S3.3.2, Based on the physical meaning of the power load pattern and the known system characteristics, construct the structure of the Bayesian network. The nodes in the network include the input variable V. k And system state nodes (such as system stability, resonance risk level, etc.). The directed edges between nodes represent the causal relationships between variables. The structural design can refer to expert knowledge or be automatically generated through structural learning algorithms.

[0111] Node definition: System state nodes represent the resonance risk level of the power system, with possible values ​​including low risk, medium risk, and high risk. Input nodes, as parent nodes of system state nodes, influence the resonance risk level of the system.

[0112] Causality: In Bayesian networks, directed edges establishing causal relationships connect input nodes to system state nodes. Different combinations of input variables will influence the values ​​of the system state nodes with a certain probability.

[0113] S3.3.3 defines a conditional probability table (CPT) for each Bayesian network node, describing the probability of different input node values ​​affecting the system state nodes. The conditional probability table can be determined through statistical analysis of historical data.

[0114] Historical data training: Using historical operating data of the power system, the frequency of the impact of different phase differences and resonance intensities on the system's resonance risk is statistically analyzed. Conditional probability tables for each node are determined using maximum likelihood estimation (MLE) or Bayesian estimation methods.

[0115] S3.3.4 After constructing the conditional probability table, the Bayesian network can perform inference analysis. After inputting the phase differences and resonance intensities of different load mode combinations, the posterior probability distribution of the system state nodes is calculated using Bayesian inference algorithms (such as belief propagation algorithms or variational inference algorithms).

[0116] Risk Level Determination: Based on the posterior probability distribution, the resonance risk level of the system is determined. Load patterns with a posterior probability greater than or equal to the high-risk threshold are classified as high-risk, indicating that these patterns may cause severe system instability or resonance effects in the current or future timeframe. Load patterns with a posterior probability not less than the high-risk threshold and greater than or equal to the low-risk threshold are classified as medium-risk, indicating that these patterns may cause some fluctuations in the system, but are not sufficient to cause major failures or instability. Finally, patterns with a posterior probability less than the low-risk threshold are classified as low-risk, indicating that these patterns have a relatively small impact on the system under current conditions and are within a safe and controllable range. This classification method ensures that targeted response strategies can be adopted for different load patterns according to their risk levels during subsequent processing, thereby optimizing the operational safety and stability of the power system.

[0117] Risk Analysis Report: Outputs comprehensive risk analysis results for each load pattern combination, including the probability distribution of each risk level and its potential impact. These results can be used for further decision support, such as dynamically adjusting system operating parameters or early warning systems.

[0118] Comprehensive risk analysis using Bayesian networks can not only quantify and identify the potential resonance risks of various load mode combinations in a power system, but also dynamically adjust and optimize the system's operating strategies. Bayesian network models effectively integrate multidimensional information from the phase difference matrix and resonance intensity matrix, using conditional probability inference to accurately predict the risk level under different load modes. This process significantly improves the system's responsiveness to complex dynamic behaviors, helping managers identify high-risk modes in advance, take preventative measures, and avoid potential system instability or failures. Ultimately, comprehensive risk analysis using Bayesian networks helps improve the accuracy and adaptability of power load forecasting models, enhances the operational stability of the power system in complex environments, and provides crucial assurance for the long-term reliable operation of the power system.

[0119] S3.4, based on the risk analysis results of the Bayesian network, the focus is on dynamically adjusting the weights of the identified high-risk load pattern combinations. A Genetic Algorithm (GA) is used to prioritize the optimization of the weights for high-risk patterns. Through selection, crossover, and mutation operations, the GA increases the sensitivity of the weights for high-risk patterns, enabling the model to have higher response speed and prediction accuracy when dealing with these potential resonance effects. The weight adjustments for medium and low-risk patterns will remain basically stable, with only minor tweaks to maintain the overall balance and robustness of the model. This approach ensures that the model can concentrate resources and computing power to prioritize high-risk patterns that may have a significant impact on system stability.

[0120] The dynamic adjustment process for load pattern weights includes the following technical features: First, high-risk load pattern combinations identified by a Bayesian network are used as input, and a genetic algorithm is employed to optimize the model's weight configuration. The selection operation in the genetic algorithm selects the best-performing weight combination for the next generation based on the performance of high-risk patterns in the fitness function. The crossover operation exchanges some genes between two weight combinations to generate new weight combinations, increasing the diversity of the search space. The mutation operation randomly adjusts some weight values ​​to prevent getting trapped in local optima and ensure global search capability. The fitness function is evaluated based on the model's prediction accuracy and response speed when dealing with high-risk patterns, and optimized generation by generation to reasonably increase the weights of high-risk patterns. The optimized weight parameters are directly updated into the model and applied in real-time to system load prediction and resonance risk assessment, ensuring the model has stronger responsiveness and prediction accuracy in high-risk scenarios. Simultaneously, by introducing an adaptive mutation rate, the intensity of the mutation operation is dynamically adjusted to balance the model's sensitivity and robustness, maintaining stability and flexibility when dealing with complex load changes.

[0121] S3.5 stores the optimized weights of high-risk load patterns and associates them with corresponding timestamps and power load scenarios for rapid retrieval in future forecasting and real-time applications. Simultaneously, a feedback mechanism for high-risk patterns is established, comparing actual operational data with the high-risk prediction results of the Bayesian network. If, in actual operation, a high-risk pattern is confirmed to have indeed triggered resonance effects or other instability phenomena, the model will further strengthen the weight adjustment for that pattern, ensuring a more sensitive and accurate response to similar situations in the future. Conversely, if some high-risk predictions do not exhibit the expected severity in reality, the model will appropriately adjust the weights of these patterns to prevent false alarms or resource waste due to oversensitivity. Through this feedback mechanism, the model can continuously optimize its predictive capabilities in actual operation, ensuring the long-term stability and reliability of the power system.

[0122] The beneficial effect of step S3 lies in achieving a comprehensive and accurate identification of nonlinear resonance effects in power systems through the combined analysis of two indicators: resonance intensity in the frequency direction and mode phase difference in the time direction. First, frequency-direction resonance intensity analysis effectively detects the energy accumulation of power load signals within specific frequency ranges, revealing potential resonance risks at these frequencies, particularly high-frequency resonance. This analysis allows the system to provide early warnings of high-energy points that may lead to instability, enabling preventative measures to be taken in advance. Second, time-direction mode phase difference analysis captures the phase synchronization and offset between different load modes in the time domain, accurately identifying the dynamic behavior of the system during critical periods, especially identifying phase shifts and coupling effects between different modes that may lead to system misalignment or resonance. This combined analysis not only expands the depth and breadth of risk identification but also achieves a comprehensive assessment of complex nonlinear risks through synergy, avoiding omissions and misjudgments that may arise from a single analytical dimension. Meanwhile, by dynamically adjusting the weights of high-risk patterns through genetic algorithms, the model can respond to potential system risks more flexibly and accurately, thereby significantly improving the adaptability, robustness, and overall prediction accuracy of the power load forecasting model, and providing a reliable guarantee for the stable operation of the power system.

[0123] In power load forecasting and optimization, the risk level of different load patterns directly affects the stability and responsiveness of the system. Traditional single forecasting models struggle to fully capture the complex dynamic behaviors under different risk levels, leading to decreased forecast accuracy and increased system risk. By classifying load patterns into high, medium, and low risk levels according to resonance risk assessment results, and utilizing deep neural networks for multi-pattern forecasting fusion, the characteristic information of various patterns can be effectively integrated, improving the forecasting model's adaptability to complex load changes. Furthermore, combining this with particle swarm optimization algorithms to adjust pattern weights in real time allows the model to flexibly respond to risk changes over different time periods, ensuring high accuracy of forecast results and high stability of system operation.

[0124] Step S4 includes the following:

[0125] S4.1, the load patterns classified into high, medium, and low risk levels are input into the Deep Neural Network (DNN), and necessary preprocessing is performed before input. Preprocessing includes normalizing the input data to eliminate dimensional differences between different patterns, and using feature selection methods to filter out the key features most influential on the prediction results, reducing input data redundancy and improving network training efficiency. For high-risk patterns, higher input weights are used to increase their influence in prediction, ensuring the model's response to these patterns is more sensitive.

[0126] S4.2, based on the preprocessed load pattern data, a deep neural network is used to perform multi-pattern prediction fusion for high, medium, and low risk levels. During training, a multi-layer network structure is set up, and the backpropagation algorithm is used to optimize the network's weights and bias parameters to minimize prediction errors. High-risk patterns, due to their significant impact on system stability, are given a higher learning rate during training to accelerate the optimization process of their weights. The network generates a unified prediction output by fusing diverse features of patterns at different risk levels, covering all possible system states.

[0127] S4.3, based on the trained deep neural network, uses Particle Swarm Optimization (PSO) to adjust the weights of the load patterns in real time. PSO simulates the movement of particles in the search space to find the optimal weight configuration. Each particle represents a possible weight combination, and the particle's fitness is determined by the prediction accuracy of the deep neural network. During the search process, particles update their velocity and position based on their own experience and the optimal position of the swarm, gradually approaching the optimal solution. For high-risk patterns, PSO dynamically adjusts the weights to ensure that the prediction accuracy of these patterns reaches its optimal level at different time periods; for medium- and low-risk patterns, PSO maintains basic accuracy while moderately reducing the weights to maintain the overall balance of the model.

[0128] In step S4.4, after weight adjustment, the deep neural network outputs the fused prediction results. To verify the model's accuracy, the prediction results are compared with actual load data to identify potential error sources. Based on the feedback, the particle swarm optimization process is fine-tuned to further improve the model's accuracy and reliability. Special attention is paid to the prediction results of high-risk modes to ensure the system can respond and issue early warnings in a timely manner under these modes. Ultimately, through this series of steps, an optimized, multi-mode fusion load prediction model is generated, possessing high accuracy, high adaptability, and flexible handling capabilities for modes with different risk levels.

[0129] By classifying load patterns into high, medium, and low risk levels in step S4 and inputting the data into a deep neural network for multi-pattern prediction fusion, the model's ability to identify and predict complex power load patterns is significantly improved. Prioritizing high-risk patterns ensures the system's response speed and accuracy in the face of potential major risks, while moderate adjustment of medium- and low-risk patterns maintains the overall balance of the model and avoids excessive resource allocation. The introduction of particle swarm optimization allows the model to dynamically adjust pattern weights based on real-time data during actual operation, ensuring that the contribution of each pattern to the prediction results remains optimal across different time periods. This not only improves the accuracy and robustness of the prediction model but also enhances the stability and security of the power system in the face of complex load changes, providing strong support for the reliable operation of the system.

[0130] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0131] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0132] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A power load forecasting and optimization method based on artificial intelligence, characterized in that, Including the following steps: S1. Wavelet transform is used to decompose historical power load data into multiple scales and extract load pattern features at each frequency level. The instantaneous amplitude and instantaneous frequency of the frequency components are obtained through Hilbert-Huang transform to generate multi-dimensional feature vectors. S2, input the multidimensional feature vector into the support vector machine, identify the nonlinear modes in the load through the kernel method; use self-organizing map to perform cluster analysis on the modes to identify potential resonant mode combinations; S3 combines two indicators—resonance intensity in the frequency direction and mode phase difference in the time direction—and uses a Bayesian network to comprehensively analyze resonance risk, classifying load patterns into high, medium, and low risk levels; and uses a genetic algorithm to dynamically adjust the weights of load patterns identified as high-risk. S4. Based on the risk classification results, the classified patterns are input into a deep neural network for multi-pattern prediction fusion, and the pattern weights are adjusted in real time through particle swarm optimization to optimize the model prediction accuracy.

2. The method for predicting and optimizing power load based on artificial intelligence according to claim 1, characterized in that: Step S1 includes the following: S1.1, Preprocess historical power load data; S1.2, the preprocessed power load data is input into the wavelet transform model, and the original load data is decomposed into frequency components of different scales using the multi-scale decomposition method to generate multiple frequency band signals, which correspond to load patterns at different time scales. S1.3, taking the wavelet decomposition results of each frequency band signal as input, and using Hilbert-Huang transform to analyze each frequency band signal; first, applying empirical mode decomposition to decompose the signal into intrinsic mode functions; then, performing Hilbert transform on each IMF to obtain the instantaneous amplitude and instantaneous frequency of the signal; S1.4, combine the instantaneous amplitude and instantaneous frequency of each frequency band with the original wavelet coefficient matrix to construct a multidimensional feature vector; S1.5 stores the constructed multidimensional feature vector matrix and associates it with the corresponding timestamp and power load scenario.

3. The method for predicting and optimizing power load based on artificial intelligence according to claim 2, characterized in that: Step S2 includes the following: S2.1, take the stored multidimensional feature vector matrix as input data and perform preprocessing; S2.2, input the preprocessed multidimensional feature vector into the SVM model for training, optimize the model's decision boundary, and enable it to distinguish different load patterns; S2.

3. After the SVM model is trained, the multidimensional feature vector is input into the trained model again to perform nonlinear pattern recognition. S2.

4. Use the nonlinear patterns identified by SVM as input to perform self-organizing map clustering analysis; S2.5 combines the cluster analysis pattern features with the identified resonance patterns and stores them, associating them with their corresponding timestamps and power load scenarios.

4. The power load forecasting and optimization method based on artificial intelligence according to claim 3, characterized in that: Step S3 includes the following: S3.1, Based on the nonlinear modes identified in step S2, the resonance intensity of each load mode in different frequency ranges is first extracted; S3.2, Based on the identification of resonance intensity, calculate the mode phase difference in the time direction; S3.3, using the phase difference matrix and the phase difference matrix as input variables, construct a Bayesian network model to determine the risk level of the nonlinear resonance effect; The resonance risk level of the system is determined based on the posterior probability distribution; Load patterns with a posterior probability greater than or equal to the high-risk threshold are classified as high-risk. Load patterns with posterior probabilities not less than the high-risk threshold and greater than or equal to the low-risk threshold are classified as medium-risk. Patterns with posterior probabilities less than the low-risk threshold are classified as low-risk. S3.4, Based on the risk analysis results of the Bayesian network, dynamically adjust the weights of the identified high-risk load pattern combinations; Genetic algorithms are used to prioritize and optimize the weights of high-risk patterns. S3.5 stores the optimized high-risk load pattern weights and associates them with the corresponding timestamps and power load scenarios.

5. The power load forecasting and optimization method based on artificial intelligence according to claim 4, characterized in that: Complete steps to obtain the resonant intensity of each load mode in different frequency ranges: S3.1.1 First, the preprocessed power load signal x(t) is converted to the frequency domain, and a Fourier Transform (FT) is used to obtain the frequency domain signal X(f), where f represents the frequency; the frequency range F is constructed as multiple discrete intervals F = {f1, f2, ..., f...} n }, where f i This represents the i-th frequency interval; by performing a Fourier transform on the power load signal, the spectral signal X(f) of each frequency interval is obtained. i ): S3.1.2, after obtaining the spectral signal for each frequency interval, calculate the instantaneous energy density for each frequency interval; construct the instantaneous energy density E(f i Let ,t) be the square form of the signal energy at time t; it is calculated using a nonlinear transformation of energy density, using the following formula: Among them, α, β, and γ are nonlinear conversion parameters used to adjust the calculation method of instantaneous energy density in order to capture the dynamic changes of the signal in different frequency ranges; S3.1.3, after calculating the instantaneous energy density for each frequency range, extract the resonance factor R(f) within the corresponding frequency range. i ): Where, θ(f) i ,t) represents the frequency range f i The instantaneous phase angle is obtained from the time-domain signal through Hilbert transform; S3.1.4, for each frequency range, the resonance factor, representing the original resonance intensity of the corresponding frequency range, is normalized; the normalized resonance intensity is denoted as S(f i ); S3.1.5 Finally, the normalized resonance intensities of each frequency range are organized into a resonance intensity matrix D(M), where each row of the matrix corresponds to a frequency range and each column corresponds to the resonance intensity of different time windows.

6. The power load forecasting and optimization method based on artificial intelligence according to claim 4, characterized in that: Complete steps to obtain mode phase differences in the time direction: S3.2.1, First, the instantaneous phase information is extracted from the power load signal x(t) using the Hilbert transform; for a given power load signal, its corresponding analytic signal z(t) is obtained through the Hilbert transform, and the analytic signal is defined as: in, The Hilbert transform of x(t) is represented, where j is the imaginary unit; the instantaneous phase θ(t) is calculated from the phase angle of the analytic signal. θ(t) represents the phase change of the load signal at each moment, laying the foundation for subsequent phase difference analysis; S3.2.2, after obtaining the instantaneous phase, calculate the phase synchronization index (PLV) to measure the degree of phase synchronization between the two load modes x1(t) and x2(t); the formula for calculating the phase synchronization index PLV is: Where N is the number of sampling points within the time window, λ1(t) k ) and θ2(t k ) represent the two modes at time t k The instantaneous phase; S3.2.3, After obtaining the phase synchronization index, calculate the phase difference metric (PLI) to describe the directionality of the phase offset; the formula for calculating the phase difference metric (PLI) is: Where sign(·) represents the sign function, which is used to preserve the directionality of the phase shift, and λ is a nonlinear extension parameter used to adjust the sensitivity of the phase difference measurement; S3.2.4 Integrate the phase difference measurements within each time window to calculate the overall phase difference intensity D(T), as shown in the formula: Where T represents the total time interval [t0, t1], t c ρ is the time center point, and ρ is the time decay factor, used to adjust the influence of different time periods on the overall phase difference intensity. S3.2.5 Finally, the overall phase difference intensity calculated at each time window will be organized into a phase difference matrix S(F), where each row of the matrix corresponds to a load mode combination and each column corresponds to the phase difference intensity at different time periods.

7. The method for predicting and optimizing power load based on artificial intelligence according to claim 6, characterized in that: Step S4 includes the following: S4.1 Input the load patterns that have been classified into high, medium and low risk levels into the deep neural network and perform preprocessing before input; S4.2, Based on the preprocessed load pattern data, use a deep neural network to perform multi-pattern prediction fusion for high, medium and low risk levels; S4.3, Based on the trained deep neural network, the particle swarm optimization algorithm is used to adjust the weights of the load pattern in real time; S4.4 After completing the weight adjustment, the deep neural network outputs the fused prediction result.

Images